Python SDK

​

Mistral AI and AWS Bedrock wrappers

• Added pandaprobe.wrappers.wrap_mistral for the official mistralai>=1.0.0 client. Instruments chat.complete / chat.complete_async (blocking) and chat.stream / chat.stream_async (streaming) — every call automatically produces an LLM span with input, output, model, model parameters, and token usage. Mistral’s native prompt_tokens / completion_tokens / total_tokens map directly onto PandaProbe’s canonical names; UNSET sentinel kwargs are stripped before logging.
• Added pandaprobe.wrappers.wrap_bedrock (beta) for the AWS bedrock-runtime client (boto3 and aioboto3). Instruments the recommended Converse API (converse, converse_stream) end-to-end — system blocks are hoisted into the universal-schema messages list, text-only content blocks are flattened, the streaming response dict ({"stream": ..., "ResponseMetadata": ...}) shape is preserved with only the inner iterator wrapped, time-to-first-token is captured on the first contentBlockDelta, and final usage (inputTokens / outputTokens / totalTokens and cache-read / cache-write counts) is read from the trailing metadata event.
• Bedrock legacy InvokeModel path is also instrumented (invoke_model, invoke_model_with_response_stream) with best-effort body parsing across Anthropic-on-Bedrock, Mistral, Llama, Titan, and Cohere body shapes.
• Async support for Bedrock is opportunistic — aioboto3 is detected at runtime via inspect.iscoroutinefunction plus client-class-module heuristics, so you can use it without listing it as a dependency.
• New optional dependencies: pip install "pandaprobe[mistral]" and pip install "pandaprobe[bedrock]".
• New examples in examples/mistral/ (chat completion, streaming, multi-turn with sessions) and examples/bedrock/ (Converse, Converse streaming, legacy InvokeModel with the Anthropic Claude body shape).

​

DeepAgents integration

• Added pandaprobe.integrations.deepagents.DeepAgentsCallbackHandler for tracing deep agents built with create_deep_agent. A single handler captures the parent agent and every sub-agent dispatched via the built-in task tool — sub-agent invocations forward callbacks / tags / configurable automatically, so the entire nested run is one trace.
• Span tree faithfully records the task (TOOL) → <subagent> (AGENT) nesting, preserving the LangChain universal schema across the dispatcher boundary.
• Trace name remap: DeepAgents wraps a LangGraph compiled graph (root reports name="LangGraph"); rewritten to "DeepAgents" while custom user-given graph names pass through.
• New optional dependency: pip install "pandaprobe[deepagents]" (requires Python ≥3.11).
• New examples in examples/deepagents/: simple agent, sub-agent dispatch via the task tool, and multi-turn with sessions.

​

Test infrastructure

• Test suite is now hermetic: a session-level fixture in tests/conftest.py pins PANDAPROBE_ENDPOINT to a fake host and clears every PANDAPROBE_* config var, so developer-machine env exports cannot leak into tests or accidentally hit production. Pinned by a regression-test class.

​

LangChain integration

• Added pandaprobe.integrations.langchain.LangChainCallbackHandler for tracing LangChain agents built with langchain.agents.create_agent and plain LCEL pipelines (prompt | model | parser).
• Refactored shared callback logic into a private pandaprobe.integrations._langchain_core package. LangGraphCallbackHandler is now a thin subclass — behavior is unchanged for existing users; future LangChain-family integrations share the same base.
• Trace name remap: create_agent’s root reports name="LangGraph", LCEL chains report name="RunnableSequence" — both rewritten to "LangChain" while custom user-given names pass through.
• Improved safe_output to faithfully serialize @dataclass instances (e.g. LangGraph Command objects) into structured JSON instead of repr() strings, preserving routing fields.
• Diagnostic-label fix: error logs in _finalize_trace now correctly reflect the active integration’s name (PandaProbe LangChain callback failed… vs PandaProbe LangGraph callback failed…).
• New optional dependency: pip install "pandaprobe[langchain]".
• New examples in examples/langchain/: simple agent, LCEL chain, multi-turn with sessions.

• Fixed installation of optional dependency extras (e.g. pandaprobe[openai], pandaprobe[gemini]) by quoting glob-pattern extras to avoid shell unmatched-glob errors on common shells.
• Minor example and documentation corrections.

• Documentation polish: page titles and copy fixes across the SDK reference.

• Updated and expanded Python SDK documentation.

​

Initial public release of the Python SDK

• Tracing core: pandaprobe.init, pandaprobe.trace, pandaprobe.span decorators, TraceContext/SpanContext context managers, background-batched transport with retry/backoff and atexit flushing.
• Provider wrappers: OpenAI, Anthropic, and Google Gemini — automatic LLM span capture without code changes.
• Framework integrations: LangGraph, Google ADK, Claude Agent SDK, CrewAI, and OpenAI Agents SDK.
• Session and user context: pandaprobe.session(...) / pandaprobe.user(...) ContextVar-based propagation across decorators, wrappers, and integrations.
• Programmatic scoring: Client.score(trace_id, name, value, ...) for trace-level evaluation signals.
• Auto-init from environment: PANDAPROBE_API_KEY + PANDAPROBE_PROJECT_NAME env vars trigger lazy client creation; PANDAPROBE_ENABLED=false disables silently.