Unify session and CI verification telemetry?

[kyle-sexton]

claude-ops instruments local sessions (OTEL DuckDB + 8 hooks); CI emits no telemetry. Boris step-1 guardrail extends to OTel into existing observability stack. Decide whether/how to unify session + CI verification telemetry. Evidence: explore-plugins-inventory.md § Verification surface 'Missing from the loop'; gap-map.md open decision 4.

[kyle-sexton]

🔒 claim: session-61a34b03

[kyle-sexton]

Resolved: YES — unify, at the contract layer. Standards-only telemetry contract + trace propagation; sink explicitly OUT of the contract.

The contract (tool/machine/repo/user/org agnostic by construction):

1. Every execution context emits standard OTLP — interactive session, CI, and the chartered autonomous runner — pinned to the OpenTelemetry CI/CD + VCS semantic conventions (Release Candidate as of Jul 2026: cicd.pipeline.run.id, cicd.pipeline.result, cicd.pipeline.task.run.result, task types, vcs.change.id, vcs.ref.head.revision). We pin upstream vocabulary, never invent parallel schema.
2. One minimal custom attribute namespace carries the work-item ID — semconv has no work-item-tracker attribute (verified gap), so this is the single custom addition. It is the Return accounting: in scope, and in what shape? #246 return-accounting join key: machine cost/lifecycle telemetry correlates to the tracker item holding the human-attested value fields.
3. Trace context propagates across the chain (W3C TRACEPARENT): trigger → CI run → agent session form one causal tree. Verified: headless agent sessions inherit TRACEPARENT natively, so a CI job or runner drain parents the agent session trace. Unification is achieved by context propagation, not by merging sinks.

Sink binding is deployment-owned (out of contract): an org with an existing observability stack points emission at it (Boris step-1 guardrail verbatim: 'OTel into existing observability stack'); no stack → free default = OTLP file artifacts + query-on-read; a running backend (self-hosted or hosted/paid) = explicit opt-in, never default. Zero new cost by default.

Fresh-eyes consequence (recorded, actioned post-graduation): the existing local pipeline is one conforming sink instance, not privileged. Native agent-CLI OTel export (cost, tokens, tool events) likely part-duplicates the local hook-envelope layer — anything reimplementing built-in telemetry gets scrapped. Audit issues to be filed via the work-item flow after map graduation, referencing this resolution.

Rejected: shared physical sink as default (new infra/cost/secret surface; fine as opt-in instance); no unification (fails the step-1 guardrail, starves the #246 join and the runner observability seam per #243/#244); mandated reference backend (binds every adopter to one sink choice, violates agnostic posture); human-readable summaries only (unqueryable, no join).

Deferred to design/architect: custom namespace naming, exact semconv version pin, CI-side emission mechanics (exporter/action choice), sink adapter shapes, native-vs-hook telemetry audit scope.

Boris-alignment verified before locking: step-1 guardrail satisfied in its literal form (standard OTel into whatever stack exists); no step-skipping (contract now, emission mechanics at build, no speculative backend); trust loop fed (verification outcomes become queryable evidence for #243 C2 promotion); 'let Claude kick off Claude' audit trail strengthened — one trace tree across trigger/CI/agent. Evidence: .work/ai-adoption-ladder/RESEARCH-telemetry-unification.md (primary-sourced, falsification pass run — 'semconv stable' claim falsified to Release Candidate).

[kyle-sexton]

Source anchors (added for durability — the design threads re-anchor to this artifact):

• Boris Cherny, "Step & your role" (Jul 16, 2026) — Google Doc (reliable copy): https://docs.google.com/document/d/1R91ayvj7uvlxgNi--__2-Bf3w8x5r1nF-xIBN7ds8Ns/edit
• Original claude.ai artifact (intermittently erroring as of 2026-07-17): https://claude.ai/code/artifact/bfdfaef9-bc62-4dfe-ba9e-c58a26c9accf