Screener capacity
Audited single-shot screening providers, trusted builds, and parked failures with manual retry control.
Provider routing
Awaiting controller r0All lanes GCE onlyHetzner handles normal work. GCE may claim separate, still-unclaimed submissions only during a primary-node outage or sustained backlog overflow. A failed lane parks its own attempt and never moves to the next provider as an automatic retry.
Capacity controller has not checked in
The independent GCE watchdog may conservatively scale out while the controller is stale and queued work exists. Hetzner node controls and the audited overflow threshold will appear again when the controller heartbeat recovers.
Infrastructure retries
Derived from screening attempt history when this page loads; nothing is stored. Capped agents, and aged-out agents (last infrastructure failure older than the maximum age; counted, not listed individually), wait for an operator retry. The breaker is per signature: reason code, provider, and lane. A breaker with a known provider holds only workers on that provider; a worker on another provider can still claim its agents by backoff alone.
0 parked · In backoff 0 · Held by breaker (that provider) 0 · Probe due 0 · Due now 0 · Capped (operator) 0 · 0 not admitted · 0 aged out (not listed; need an operator) · 0 open and 0 half-open of 0 breakers · as of Nov 3, 11:12 PM UTC
- Backoff
- 10 min doubling to 1 h, ±20% jitter
- Stops after
- 8 consecutive failures or 24 h since the last failure
- Retries for l2-late-result
- 1 automatic, then operator
- Breaker trips at
- 3 distinct agents within 5 min
- Breaker open / probe
- open 10 min, then one probe per 5 min
- History read
- 48 h
- Retried codes
- docker-build-infrastructure, worker-claim-not-started, l2-runtime-evidence-unavailable, source-review-adjudicator-key-unavailable, image-upload-platform-unavailable, l2-late-result
Circuit breakers
No infrastructure failures in the history window.
Parked agents
No agent is parked on an infrastructure failure.