You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Problem — The cloud runner exposes no observability surface: probes of /health, /status, /instances, /gpus, /quota return empty bodies; the GPU SKU is unknowable before the first launch (a parallel session's finding, 2026-08-18 — "A100 unverified; vault notes suggest H100 80GB for GKP-class"); receipts carry no GPU accounting, so spend can't be gated or attributed. The GPU-plane spec (vault spec-20260818-gpu-plane) records this as the runner contract — the durable ask that outlives today's DNS incident.
The contract (cloud-side)
GET /health → liveness + version · GET /quota → remaining budget/wallclock · GET /instances → live jobs (id, sku, age)
Important
Problem — The cloud runner exposes no observability surface: probes of /health, /status, /instances, /gpus, /quota return empty bodies; the GPU SKU is unknowable before the first launch (a parallel session's finding, 2026-08-18 — "A100 unverified; vault notes suggest H100 80GB for GKP-class"); receipts carry no GPU accounting, so spend can't be gated or attributed. The GPU-plane spec (vault
spec-20260818-gpu-plane) records this as the runner contract — the durable ask that outlives today's DNS incident.The contract (cloud-side)
GET /health→ liveness + version ·GET /quota→ remaining budget/wallclock ·GET /instances→ live jobs (id, sku, age)gpu_sku,gpu_seconds,cost_usd(when billable)max_wallclock(client now ALWAYS sends it — Every remote launch carries a wall-clock cap (#GPU-plane anti-hang) #420) and reports cap-hit as a DISTINCT terminal, not a generic failureClient side is ready to consume: FakeCloud fixtures extend in minutes; first consumer is the receipt-accounting issue. Cloud-side owned (Jack).