EventHorizon item/effect SVG generation now requests direct GLM-5 output with thinking disabled and a 2,048-token output limit. This applies only to icon requests; gameplay and editor inference payloads are unchanged.
The automatic worker scan interval changed from 60 seconds to 15 seconds. Each pass still permits at most one model-backed generation, all pods still share one cluster-wide icon-generation lease, and generation continues to defer around foreground turns. A pass may propagate up to 128 existing template icons to matching session items without consuming model-generation capacity.
Before the change, three production icon calls took 22.6-99.3 seconds (59.3 seconds average) and used 3,561-11,312 completion tokens (6,797 average) despite returning only 474-1,122 bytes of SVG. GLM-5 reasoning was the dominant latency.
76e7e62cf3b8d64b4df901f611b8acf70764c8ad4723 backend deploy and 4724 frontend verification, both successful/healthz OK; /readyz reports PostgreSQL and Qdrant readyOne explicitly approved, synthetic, non-persisting effect-icon request was made on 2026-07-29. It did not alter any stored icon, item, effect, session, or game state.
2001.795 secondsllm_usage_events: 1,764 ms120 / 238 / 358200Compared with the three-call pre-change average, the verified provider duration is approximately 33.6 times faster and completion-token use is 96.5% lower.
The reuse path was also observed propagating 39 existing template icons to session overrides between 15:58:11.120722Z and 15:58:11.673910Z: about 0.55 seconds total and no model calls.
Revert merge 76e7e62 through the normal repository deployment workflow. No schema or stored icon format changed.