Immutable
release. Only release title and notes can be modified.
What's Changed
- Add Russian localization (#1782) by @DrMaks22
- Add safe custom JSON Agent Channel runner (#1652) by @mimeding
- Add Agent Channel inbox audit workbench (#1784) by @mimeding
- Repin vmlx-swift to c976fcbd — Ornith prefill counter advances (vmlx #105) (#1779) by @jjang-ai
🐛 Bug Fixes
- gated Rampart PII scan and media prep as exclusive GPU producers (#1799) by @RaajeevChandran
- Disconnect must not evict the BatchEngine (cold-load hangup crashed the server) (#1796) by @jjang-ai
- Generation params: wire seed into the sampler, accept string stop, native penalties, min_p (#1793) by @jjang-ai
- triage and fix sentry app hangs and crashes (#1781) by @RaajeevChandran
- Repin vmlx-swift b6eda04f — MLX Metal stream-map thread-safety (fixes concurrent-GPU EXC_BAD_ACCESS crashes) (#1780) by @jjang-ai
🧰 Maintenance
- Model catalog IA cleanup: family-grouped cards, variant picker, chat-only org fetch (#1800) by @tpae
- Cut per-turn prompt cost: remote prompt caching + local KV prefix reuse + measured text trims (#1798) by @tpae
- Subagent harness hardening: admission gate, mid-run interrupt, tool-capable workers, RAM coexistence, residency crash lane (#1797) by @tpae
- Repin vmlx-swift: deterministic RMSNorm shift + Mistral VLM full-resolution images (#1791) by @jjang-ai
- Settings IA cleanup: grouped sidebar, relocations, unified card primitives (#1792) by @tpae
- Upgrade evals harness: repeat/resume runs, judge audit + calibration, micro-perf and prompt-injection suites (#1789) by @tpae
Full Changelog: 0.21.3...0.21.4