App telemetry vs agent observability
Two features share English words. They are different products.
| Agent-run observability | App telemetry | |
|---|---|---|
| CLI / flag | pipeline-kit obs … | pipeline-kit features enable telemetry or RUN_TELEMETRY: true |
| Specialist | none (hooks + scorer) | telemetry-agent |
| Question it answers | Did the agent waste tools, miss the allowlist, skip verify? | Which product events does this feature owe, if any? |
| Output | JSONL ledger, Langfuse traces and scores | features/{slug}/telemetry-contract.md |
| Default | Off until obs install | Off. 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.