THE FIELD GUIDE / AI & THROUGHPUT
Parallel triage, timing & retries
Reduce queue delay without weakening the model, evidence or review boundaries.
Keep the full investigation
Parallel work retains the configured model and reasoning effort, authorized input, conditional SSH evidence checks, second-stage requests, structured validation, explanations and uncertainty. An unavailable or incomplete assessment stays unresolved. Faster queue handling never means silently choosing a weaker model or dropping evidence.
Configure concurrency separately
Use bounded, configurable threaded execution for independent investigations, with a supported serial setting. Separate per-item claims and renewable leases, worker/boot ownership, independent database sessions and explicit policy context protect concurrent work. An administrator sets one to four investigation slots per server, with two as the default. Lowering the limit preserves live leases and blocks new admission until occupancy falls below it. Global capacity, lifecycle and provider controls still apply. Conflicting target operations remain serialized.
Keep the queue fair
Investigations, scoring and other AI purposes use fair admission, priority/aging and bounded per-host occupancy. Slow scoring or one busy host cannot take every slot. Work waiting for rate refill, a daily reset or retry eligibility is deferred durably instead of sleeping in an occupied execution thread. Pause and sharing state are rechecked before each provider call.
Follow each stage
AI activity shows queue age, admission, provider transport, conditional SSH evidence, receipt/result persistence and tracked run timing alongside calls, tokens and estimated cost. Inspect active/queued capacity, oldest wait and the next eligible time. Connection pooling reuses transport safely; it is separate from the number of investigations running.
Retry the right failures
Transient network and provider overload failures use bounded exponential backoff with jitter and applicable Retry-After floors. Authentication, billing, invalid schema/model configuration and invalid or incomplete results remain review-blocked until corrected or explicitly retried. Attempts, classification and eligibility persist across process restarts. Manual retry respects provider delay, pause, spending policy and current evidence.
Recover ownership honestly
Lease/heartbeat-aware reconciliation handles lost workers, shutdown and interrupted follow-up stages without accepting a stale owner’s result. Drain stops new work and preserves recovery state. A lost provider response can still be billable; retained receipts and uncertain charges are not converted to zero or treated as exactly-once paid execution.
Open the application reference for capability scope and technical details.
Open Open SECURITY_IMPLEMENTATION.md on GitHub ↗