Python SDK

​

LangChain

• Root output: Reasoning content blocks become the answer string on the root trace output. The generation span keeps the full block list. Tool-call payloads stay unchanged.
• Empty results: A successful non-root chain or tool that returns None ends with {"result": "none"}. Root traces and errors are unchanged.

​

LangChain

• Idle traces: An open trace with no activity for open_trace_ttl (default 2 hours) is finalized and sent. A long-running loop is not closed mid-run. The sweep runs from chain, model, and tool starts, at most once per eviction interval.

​

Tracing

• Before send: before_send can redact or drop an ingest payload without a custom HTTP transport.
• Payload cap: max_payload_bytes caps the payload. deduplicate_span_content shrinks repeated history before the cap is applied.

​

LangChain

• Span filter: exclude_span_names and include_span drop spans, including noisy LangGraph middleware, while children stay attached to the nearest kept ancestor.
• Model name: When a chat model is not serializable, the handler reads ls_model_name from run metadata after invocation params.

​

Ingest

• User-Agent: A custom transport sends uselemma-tracing/<version> unless you set User-Agent yourself.

​

LangChain

• LangSmith middleware: When LangSmith wraps LangChain middleware and skips on_chain_start, nested callbacks stay on the active turn instead of becoming extra root traces.

​

Tool calls

• Failure payloads: A tool result with success: false or status of error or failed is recorded as an error. The span uses structuredContent.error when that field is set.

​

LangChain

• Callback flags: The handler subclasses LangChain’s callback base when langchain_core is installed, and sets run_inline plus the other 1.x flags. LangChain 1.6 can call the handler without crashing before a span is recorded.

​

Generations

• Token counts: Non-finite numbers, including infinity, are dropped from numeric usage fields.

​

Tracing

• One turn, two processes: start_turn, attach_turn, and apply record a host and a sandbox as one trace. The child keeps a journal and does not call ingest. The host sends once. turn.end() still raises on failure.

​

Ingest

• Fail open: trace(), async_trace(), and TraceHandle.end() catch HTTP and network errors, log them, and do not rethrow. ingest() still raises. A delivery failure does not replace your agent’s error.

​

Generations

• Model identity: A typed generation span copies the model from the invocation, the response, or response metadata when params omit it. The model is omitted on end when it is still unset. Completion text is not used as a model id.

​

Tool calls

• Encoded failures: A tool payload that encodes a failure is stored as a span error.

​

Generations

• Token usage: LangChain and OpenAI Agents record token usage, including cache and reasoning tokens from Agents SDK details.
• Provider: Generation spans set the provider system from the model provider.

​

Tracing

• Release: Pass release to the client, or set LEMMA_RELEASE, to stamp the running revision on each trace.
• Provenance: Spans record the SDK language and integration so missing usage is distinct from a real zero.

​

Tracing

• Inputs and outputs: Prompts, tool inputs, tool outputs, model output, and error messages are always recorded. The record_inputs and record_outputs arguments are removed. Passing them raises TypeError. Redact sensitive values before they reach the agent.

​

Tracing

• Error messages: Empty and missing failure text uses the same message in the Python and TypeScript SDKs. Serialized payloads use compact JSON.

​

Ingest

• User-Agent: SDK HTTP requests send uselemma-tracing/<version>.

​

Tool calls

• User-facing text: Set user_facing_message when a tool sends text directly to the person. Lemma shows that text as the message they received.

​

OpenAI Agents

• Trace contract: The processor records current-turn input and output, terminal errors, and span-bounded times. Soft tool errors are included.

​

LangChain

• Trace contract: Owned runs keep current-turn input, output, and errors. Tool calls stay nested under one root until the final answer. LangGraph uses the LangChain adapter.

​

Tool calls

• MCP errors: A tool result with isError is recorded as an error, without an invented output.
• Child timing: A tool span no longer ends after its parent generation.

​

Ingest

• One delivery: The replace option is removed. Each execution sends one complete trace. Patching a trace over time is not supported.

​

Tracing

• Environment field: The undocumented environment client field is removed.

​

Debug

• Ingest status: LEMMA_DEBUG_VERIFY checks GET /traces/ingest-status for up to 15 seconds, so a backlog does not look like a missed delivery.
• Debug flag: LEMMA_DEBUG=1 is the preferred value. true still works.

​

Debug

• Delivery logs: Debug mode logs client config and ingest responses, including failure hints.
• Smoke test: Debug verification checks that a trace reached Lemma.

​

Ingest

• Send on end: Trace handles and the OpenAI Agents processor send a trace once, when it ends. They no longer send partial snapshots along the way.

​

Tracing

• First public release: uselemma-tracing sends traces to Lemma, with debug mode and experiment mode.

​

LangChain

• Callback handler: A LangChain callback handler records chains, model calls, and tools.

​

OpenAI Agents

• Processor: An OpenAI Agents processor records agent runs.