Jev efficiency and reliability
- Share decision context and batch related questions to reduce repeated payloads.
- Remove redundant automatic hints and ground change reviews in executed edits.
- Limit review-driven effort escalation to one step per turn, applied to the next model call only.
- Ground tool, memory and document selection in the actual request, with conservative context-preservation checks.
- Record per-round model usage and validate cheap compression before replacing output.
- Keep transient Jev failures out of the status line; continue normally and allow later attempts, while retaining diagnostic logs.
Release assets are uploaded by the Release workflow after the platform builds succeed. The release becomes Latest after those assets are available.
Net cost savings and model-quality equivalence still require measurement on real workloads; no fixed savings percentage is claimed.