Python Telemetry v3.3.0

Changed

• Haystack, LlamaIndex, and CrewAI now use the OpenInference instrumentors instead of the OpenLLMetry/Traceloop ones (opentelemetry-instrumentation-{haystack,llamaindex,crewai}). The OpenLLMetry instrumentors produced no/incomplete conversations and no model/token/cost (Haystack emitted dict-key "content", LlamaIndex emitted no LLM span at all, CrewAI dumped agent-config templates and leaked the provider API key). The OpenInference instrumentors capture the real exchange: Haystack and LlamaIndex now resolve provider/model and full tool-call conversations; CrewAI captures the whole conversation (incl. tool calls/results) and no longer leaks the API key. Instrumenting is unchanged ({"haystack": haystack}, {"llamaindex": llama_index}, {"crewai": crewai}). New deps: openinference-instrumentation-haystack==0.1.34, openinference-instrumentation-llama-index==4.4.2, openinference-instrumentation-crewai==1.1.9.
• DSPy is now instrumented through litellm instead of the OpenInference DSPy instrumentor, which emitted no token usage (Arize/phoenix#3119, closed "not planned") and duplicated the conversation. DSPy runs every LM call through litellm, so instrumentations={"litellm": litellm} captures the model call with provider/model/tokens/cost and the full conversation. Breaking: the dspy instrumentation key is removed — passing {"dspy": ...} now raises (use {"litellm": litellm}). (Note: DSPy's response cache is on by default; cached calls don't hit litellm, so set cache=False on dspy.LM to record usage on repeated runs.)
• LiteLLM now uses LiteLLM's native OpenTelemetry callback instead of the OpenInference openinference-instrumentation-litellm. The native callback emits standard gen_ai.* semantic-convention spans, so LiteLLM traces now resolve provider, cost, and token usage (previously no provider was emitted → cost was always 0). Instrumenting is unchanged — still instrumentations={"litellm": litellm}. The SDK configures the callback through env-backed settings, set with os.environ.setdefault so you can override any of them: OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental (structured gen_ai.{input,output}.messages), OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=SPAN_ONLY (record prompt/response content), and USE_OTEL_LITELLM_REQUEST_SPAN=true (emit one child span per completion, so multiple calls in a capture no longer collapse onto the parent). Requires litellm >= 1.88 for the structured semconv + per-call spans; older versions still export but fall back to legacy attributes (no provider/cost).
• OpenAI now uses the OpenInference openai instrumentor (openinference-instrumentation-openai) instead of OpenLLMetry's opentelemetry-instrumentation-openai. The OpenLLMetry instrumentor emitted no llm.request.type on Responses API spans (so they resolved operation=unspecified and dropped out of the token/cost/conversation rollup) and flattened Responses tool turns into plain text. OpenInference resolves the operation natively for both Chat Completions and the Responses API, captures the full tool_call/tool_call_response turns, and records streaming token usage — provider/model/cost preserved. Instrumenting is unchanged ({"openai": openai}); Azure OpenAI (same key) benefits too.
• Google Gemini now uses the OpenInference google-genai instrumentor (openinference-instrumentation-google-genai) instead of the OpenLLMetry opentelemetry-instrumentation-google-generativeai. The OpenLLMetry instrumentor only str()-ified the current google-genai SDK's Part objects, so tool calls and a binary thought_signature leaked into message text as unreadable Python reprs. The OpenInference instrumentor structures tool calls (tool_call parts) and additionally captures reasoning-token usage. Instrumenting is unchanged — still instrumentations={"google_generativeai": genai} with from google import genai; install the google-genai SDK.

Removed

• Dropped the OpenLLMetry instrumentor deps opentelemetry-instrumentation-haystack, opentelemetry-instrumentation-llamaindex, and opentelemetry-instrumentation-crewai (replaced by the OpenInference instrumentors).
• Dropped the opentelemetry-instrumentation-google-generativeai dependency (replaced by openinference-instrumentation-google-genai).
• Dropped the opentelemetry-instrumentation-openai dependency (replaced by openinference-instrumentation-openai).
• Moved openai-agents out of runtime dependencies into the dev group — it's the host SDK for the openai-agents integration (the user installs it when they use it), not something the SDK imports, so it no longer bloats every install.
• Dropped the openinference-instrumentation-dspy dependency and the dspy instrumentation key (DSPy is observed via litellm — see above).
• Dropped the openinference-instrumentation-litellm dependency (replaced by LiteLLM's built-in OpenTelemetry callback).