mirror of
https://github.com/stablyai/orca.git
synced 2026-09-29 16:02:50 +00:00
Neil
63aee7f1ee
The inspection rate limiter counted panes when it should have counted host observations. `MAX_INSPECTION_STARTS_PER_SECOND = 8` is global, and it was spent one pane at a time, so N due panes meant an effective per-pane period of max(tier, N/8 seconds) — ~37.5s at 300 panes for a pane the code polls at 750ms. Agent-completion latency degraded monotonically as panes were added. Every local pane's inspection resolves out of the same TTL-and-in-flight- deduped process-table capture, so a whole round of them is one host observation. The queue now drains all shared-observation tasks as one round on one start, launched in a single tick. Remote panes each cost their own execution-host round trip and stay admitted one at a time. Both the budget and the cadence tiers are numerically unchanged. Disposed tasks are also compacted out in one pass instead of a splice per drop, so the per-round predicate cost is linear rather than quadratic at pane scale. No IPC, preload, wire, or main-process change: each pane keeps its existing per-pane `pty:inspectProcess` invoke.