Skip to content

Release 1.6.0

Choose a tag to compare

@github-actions github-actions released this 14 Oct 16:17
Immutable release. Only release title and notes can be modified.
6be1827

Added

  • observability-builder: OneUptime's AI / LLM views as an LLM-monitoring destination. genai.content.destination (langfuse or ops) sends prompts and answers to one tool per service, never both, and backends/oneuptime.md describes what the views read: call kinds, problems, conversations, usage and pricing.
  • observability-builder: who spent what. The spender, the team and the conversation come from the app's own verified context, decided per entry point; on the ops path AI spans carry user.id, team.id and gen_ai.conversation.id, where OneUptime attributes spend.
  • observability-builder: MCP servers and clients (workloads/mcp.md): a span per message on each side, context through params._meta, the JSON-RPC status rules, one span per tool call, and the mcp_server_operation_duration_seconds and mcp_client_operation_duration_seconds metrics.
  • observability-builder: the genai-langfuse-two-senders audit rule; validate checks for the content destination; the receiver's --genai-content-destination flag and summary lines for who spent what, AI keys on other spans, retrieval steps, MCP and Langfuse span senders.

Changed

  • observability-builder: retrieval steps carry gen_ai.operation.name=retrieval and gen_ai.retrieval.top_k; memory, planning and agent-creation step kinds are added; a sub-agent call is one invoke_agent span; the turn's input and output go on the Langfuse trace; no cost is written unless the provider reports one.
  • observability-builder: the ops allow-list keeps the MCP keys, the OpenInference and OpenLLMetry model and usage keys, and user.id and team.id; spans that aren't AI calls lose their gen_ai.*, llm.* and traceloop.* keys, which OneUptime would count as AI calls. The collector's GenAI routing does the same.
  • observability-builder: one span sender to Langfuse per service is a stated rule; all three Langfuse modes stay, with hybrid the default.
  • observability-builder: SDK notes for openai 3 and mcp 2 over httpx2, the MCP Python SDK 2.x client spans and protocol probe, OTLP exporter errors fed back into the log export, ExceptionGroup on FastAPI SERVER spans and the JavaScript exporters' warning on JSON replies. Direct export to OneUptime: ops metrics need a scraper, and the yml adds deployment.environment for exceptions. Path exclusions are decided when the response ends.
  • README and COMPATIBILITY list the 1.6.0 release.

1.6.0

Required

No migration required. genai.content.destination defaults to langfuse, the 1.5.0 behaviour,
so existing observability.yml files validate and behave unchanged.

Recommended

  • Re-run observability instrument on services that call a model, to add the user, team and
    conversation to their AI spans (OneUptime's Usage shows their spend as Unattributed without
    them), the retrieval operation, the new allow-list keys, and the removal of AI keys from spans
    that aren't AI calls.
  • Where prompts and answers should be read in OneUptime instead of Langfuse, set
    genai.content.destination: ops, with Langfuse mode hybrid or otel, or Langfuse off.
  • Instrument MCP servers and clients with references/workloads/mcp.md, including the server
    and client MCP metrics.
  • Collectors with GenAI routing: update the ops branch from references/collector/README.md
    (identity on AI spans, AI keys off other spans, the new allow-list) and list the services whose
    content destination is ops in both branches (^$ while there are none).
  • Services exporting straight to OneUptime: add deployment.environment: "${OBS_ENVIRONMENT}"
    under resource.attributes, arrange a scraper for /metrics, and add OneUptime prices for
    self-hosted or aliased models.

Verify

  • python3 -m unittest discover -s observability-builder/scripts/tests passes, and the receiver
    tests pass with uv run --no-project --with grpcio --with opentelemetry-proto --with protobuf --with pyyaml python -m unittest discover -s scripts/tests -p "test_receive.py" from
    observability-builder/.
  • A re-instrumented AI service's receiver summary shows the user, team and conversation on its AI
    calls and no AI KEYS ON OTHER SPANS alarm.