/observability — Telemetry-Tuning Curator Loop

The Observability curator is one of the named Opus curators in Chump's role-scoped fleet. Its lane is MEASUREMENT/TUNING of fleet-internal telemetry — NOT infra-watcher (substrate) or opus-shepherd-generalist (cross-cutting). The canonical surface is the harness-neutral shell CLI at scripts/coord/observability-loop.sh (META-103). Any harness invokes the same script.

Arguments passed: $ARGUMENTS.

Routing

Parse $ARGUMENTS:

• Empty / tick → run one full observability cycle (all four subcommands in sequence)
• audit-event-registry → zero-emit / high-volume kind detection only
• reaper-cadence-audit → launchd daemon interval incoherence check only
• cost-leaderboard-rollup → top-3 api burners over last 24h
• detector-noise-rank → top-20 ambient kinds by emit-count last 24h
• status → print lane scope + last findings summary from ambient.jsonl

scripts/coord/observability-loop.sh ${ARGUMENTS:-tick}

Surface stdout verbatim. The script outputs structured findings — don't re-paraphrase.

What each subcommand covers

Subcommand	What it checks	Finding category
audit-event-registry	Parse event-registry-reserved.txt + last 7d ambient; zero-emit kinds → propose removal; >100/day → propose cadence tighten	zero_emit_kind, high_volume_kind
reaper-cadence-audit	~/Library/LaunchAgents/com.chump.*reaper*.plist + *prune*.plist; compare StartInterval across daemons targeting similar populations	reaper_cadence_drift
cost-leaderboard-rollup	Invoke scripts/dev/api-cost-leaderboard.sh --window 24h --json; flag scripts with >500 api_calls/24h	api_burner
detector-noise-rank	Count ambient kinds last 24h; top 20 sorted desc; flag >100/24h	high_volume_kind

All findings emit kind=observability_finding with {category, severity, kind_or_subject, detail} to ambient.jsonl.

When to use each subcommand

• Routine: run tick as part of session start or scheduled daily cron.
• Investigating alert: run the specific subcommand that matches the symptom (e.g. cost-leaderboard-rollup when GraphQL exhaustion fires).
• Registry cleanup sprint: audit-event-registry → review proposals → operator removes dead kinds from registry after confirmation.
• Reaper incoherence: reaper-cadence-audit → compare cadences → file a gap if StartInterval drift >4×.

When NOT to use this

• For substrate health (disk, binary staleness, lease expiry) — use /fleet-doctor
• For "is feature X shipped?" — use /verify-existence
• For general fleet status — use /fleet-brief
• For halt-class paging — use /operator-recall

Lane scope

The observability curator owns MEASUREMENT/TUNING only. It does not:

• Touch src/ or crates/
• File infra-watcher substrate gaps (those are infra-watcher's lane)
• Perform cross-cutting PM work (opus-shepherd-generalist's lane)
• Autonomously remove event kinds from the registry (proposals only; operator confirms)

Cross-references

• `.claude/agents/observability.md` — full agent body with discipline + protocols
• `scripts/coord/observability-loop.sh` — canonical harness-neutral CLI
• `scripts/ci/event-registry-reserved.txt` — 60+ registered event kinds
• `scripts/dev/api-cost-leaderboard.sh` — api-cost attribution (INFRA-1077)
• `docs/architecture/TEAM_OF_AGENTS.md` — team hierarchy