Observatory

Live observability lens: fleet, models, personas, status, and history. Coggy-informed reasoning.

observatory.raindesk.dev ledger trace status free stack

Scope

fleet health — 7 machines (4 contabo, 3 vast.ai GPU)
agent status — – agents across agentslack
endpoint surface — services, APIs, reverse proxies
tmux topology — – sessions, – claude
history + anomalies — disk pressure, idle GPU burn, quota

Apparatus

metrics registry — CPU, MEM, DISK, GPU, NET per machine
time-series slices — sparkline histories, rolling 60s windows
event stream joins — agentslack SSE + tmux capture + system probes
vast.ai API — instance cost, uptime, utilization
SSH probes — contabo fleet health checks

Fleet

Ops Orchestrator

Walk the loop step-by-step and watch status shift across deploy, sweep, and quota pressure.

1Compose check

2Fleet probe

3Live sweep

4Quota pressure

Last run: never · Sequence: 0

Live Sweep

Run a live sweep against known surfaces and study the latest observation log.

Last sweep: never · Results: 0

Endpoint	Status	ms	Note

Techniques

rolling windows — 60s sparkline, 5m aggregate
alert thresholds — disk >90%, GPU idle >30m, cost >$1/hr
persona overlays — agent sphere assignments + DM routing
semantic clustering — group by domain optic

Coggy Reasoning Trace

⊢ awaiting observation data

Study Log

Chronicle of observed runs. Written by the ops loop, rendered here.

–

No observations yet. Run a sweep to begin.

health: – budget: – agents: – sweep: –