SigNoz is an open-source, OpenTelemetry-native observability platform for your team and their AI agents. Get logs, metrics, and traces in one tool with features like APM, distributed tracing, log management, infra monitoring, etc. Combined with SigNoz MCP and a native AI teammate (in SigNoz Cloud) it helps you build more resilient apps.
Span Mapper
Where each source key in the shipped definitions under pkg/modules/spanmapper/implspanmapper/fs/definitions/ comes from. A group runs on spans with an attribute key containing one of its condition substrings; sources are tried by priority and the first present wins; an existing target is never overwritten. Bump version on every change.
gen_ai.llm
Condition: model, llm.input_messages (OpenInference sets llm.model_name from the response, so a failed call has no model key).
| Source | Emitted by |
|---|---|
llm.model_name, llm.provider, llm.system, llm.token_count.prompt, llm.token_count.completion, llm.token_count.prompt_details.cache_read, llm.token_count.prompt_details.cache_write, llm.token_count.completion_details.reasoning, llm.finish_reason, input.value, output.value, embedding.model_name, reranker.model_name, openinference.span.kind |
OpenInference |
llm.request.model, llm.response.model, llm.vendor, llm.usage.prompt_tokens, llm.usage.completion_tokens, llm.response.finish_reason, llm.response.stop_reason |
OpenLLMetry before the gen_ai rename |
gen_ai.system, gen_ai.usage.prompt_tokens, gen_ai.usage.completion_tokens, gen_ai.usage.cache_read_input_tokens, gen_ai.usage.cache_creation_input_tokens, gen_ai.usage.reasoning_tokens, gen_ai.prompt, gen_ai.completion, llm.request.type, traceloop.span.kind |
OpenLLMetry, LiteLLM, CrewAI, Mistral, LangSmith, OpenRouter gateway |
gen_ai.usage.cache_write.input_tokens |
OTel semconv draft spelling, Codex |
gen_ai.response.finish_reason, gen_ai.usage.prompt_token_details.cached_read_tokens, gen_ai.usage.completion_token_details.reasoning_tokens |
Bifrost gateway |
gen_ai.usage.input_tokens.cached |
OpenRouter gateway |
gen_ai.response.model, last source for the request model |
vLLM, which writes only the response model |
ai.model.id, ai.model.provider, ai.response.model, ai.response.finishReason, ai.prompt.messages, ai.response.text, ai.response.toolCalls, ai.usage.promptTokens, ai.usage.completionTokens |
Vercel AI SDK v4, Mastra |
ai.usage.inputTokens, ai.usage.outputTokens, ai.usage.cachedInputTokens, ai.usage.reasoningTokens |
Vercel AI SDK v5 |
langfuse.observation.model.name, langfuse.observation.input, langfuse.observation.output, langfuse.observation.type, langfuse.session.id |
Langfuse SDK v3 (Python) and v4 (JS), LiteLLM langfuse_otel callback |
session.id |
OpenInference, Langfuse, Claude Code |
gen_ai.client.session_id |
opentelemetry-hooks runner (Cursor, Copilot, Claude Code hooks) |
model, input_tokens, output_tokens, cache_read_tokens, cache_creation_tokens |
Claude Code claude_code.llm_request spans (docs) |
gen_ai.tool
Condition: tool.name (OpenInference), ai.toolCall. (Vercel, Mastra), tool_call_args, tool_response (Google ADK), tool_use_id (Claude Code, opentelemetry-hooks). Not tool_name: Codex records it on the model call span.
| Source | Emitted by |
|---|---|
tool.name, tool.id, tool.description, input.value, output.value, openinference.span.kind |
OpenInference |
ai.toolCall.name, ai.toolCall.id, ai.toolCall.args, ai.toolCall.result |
Vercel AI SDK, Mastra |
traceloop.entity.input, traceloop.entity.output, traceloop.span.kind, llm.request.type |
OpenLLMetry |
gcp.vertex.agent.tool_call_args, gcp.vertex.agent.tool_response |
Google ADK |
tool_name |
Claude Code claude_code.tool spans |
gen_ai.client.tool_name, gen_ai.client.tool_use_id |
opentelemetry-hooks runner (Cursor, Copilot, Claude Code hooks) |
gen_ai.agent
Condition: agent.
| Source | Emitted by |
|---|---|
agent.name, agent.id, agent.description, openinference.span.kind |
OpenInference; agent.id and agent.description seen in no capture |
agent_name, final_result |
Pydantic AI |
crewai.agent.role, crewai.agent.id |
CrewAI |
traceloop.span.kind, llm.request.type |
OpenLLMetry |
agent_id |
Claude Code |
Coverage of documented integrations
Every integration under signoz.io/docs/llm-observability was checked by looking at the instrumentation its page installs and the keys that instrumentation writes.
- Most pages install OpenInference or OpenLLMetry, or an SDK that writes OTel GenAI semconv itself (OpenAI, OpenAI Agents, Pydantic AI, AutoGen, Microsoft Agent Framework, Semantic Kernel, Copilot Chat, LiteLLM, OpenRouter). Their keys are the rows above.
- Three write their own keys and have their own rows: Claude Code and Claude Agent SDK, Cursor through opentelemetry-hooks, vLLM (response model only).
- Codex cannot be priced by renaming:
modelsits on the sampling function spans,gen_ai.usage.*on the childhandle_responsesspan,conversation.idon thecodex.samplingphase span, never together on one span.
To check a new integration, capture a trace and look for a model, usage, tool or messages key that no row above lists.
Left unmapped on purpose
LiteLLM llm.openai.messages and llm.openai.choices (debug echo of the parent span's messages, Python repr). Total tokens and vendor cost keys (no target; the pricing processor owns cost). ai.response.msToFirstChunk and gen_ai.latency.time_to_first_token (milliseconds and seconds under one target).