Skip to main content

App telemetry vs agent observability

Two features share English words. They are different products.

Agent-run observabilityApp telemetry
CLI / flagpipeline-kit obs …pipeline-kit features enable telemetry or RUN_TELEMETRY: true
Specialistnone (hooks + scorer)telemetry-agent
Question it answersDid the agent waste tools, miss the allowlist, skip verify?Which product events does this feature owe, if any?
OutputJSONL ledger, Langfuse traces and scoresfeatures/{slug}/telemetry-contract.md
DefaultOff until obs installOff. Feature waves stub EVENTS: none

App telemetry rules

Events are extracted from the spec and gated — not brainstormed.

  • Candidates only from goals, journey endings, named errors, Must FRs, ACs, and a section that names a consumer.
  • Reliability gates G1–G7. Cap ≤ 5 events.
  • Zero passing gates → EVENTS: none (that is SUCCESS).
  • Do not emit properties off the allowlist.
  • Do not add a vendor pixel unless a Must FR names the vendor.
  • Do not use user id as a metric label. Do not log emails, passwords, or secrets.

Micro/minor always stub telemetry. Default feature waves skip telemetry-agent unless you opt in.

Wrong events are worse than none.