# Changelog

## Score

- CAI 37.2 → 38.6 (+1.4)
- Rubric changed (rubric-2026.11.2 → rubric-2026.11.4) — scores are not directly comparable.

## Lenses

- Code Health 23.8 → 23.9 (+0.1)
- Architecture 90.6 → 91.4 (+0.8)
- Maturity 63.2 → 70.3 (+7.1)
- Readiness 36.5 → 38.2 (+1.7)
- Security 63.5 → 72.0 (+8.5)
- Domain Modelling 100.0 → 100.0 (0.0)
- Accessibility 42.5 → 43.0 (+0.5)
- Performance 100.0 → 100.0 (0.0)

## Resolved (149)

- Change-coupling hub: dashboard.js → ModelSettingsScreen.swift, PerformanceScreen.swift, ModelSettingsScreenVM.swift, GlobalSettingsDTO.swift, ModelsDTO.swift, OMLXClient.swift, accuracy_benchmark.py, routes.py, config.py, settings.py (apps/omlx-web/omlx_web/static/js/dashboard.js)
- Duplicated block (10 lines × 13) (apps/omlx-mac/Sources/AppView/Screens/AppearanceScreen.swift)
- Duplicated block (10 lines × 2) (apps/omlx-mac/Sources/AppView/Screens/AskAIScreen.swift)
- Duplicated block (10 lines × 2) (apps/omlx-mac/Sources/AppView/Screens/ContextBenchScreen.swift)
- Duplicated block (10 lines × 2) (apps/omlx-mac/Sources/AppView/Screens/IntegrationsScreen.swift)
- Duplicated block (10 lines × 2) (apps/omlx-mac/Sources/AppView/Screens/SecurityScreen.swift)
- Duplicated block (10 lines × 2) (omlx/cluster/identity.py)
- Duplicated block (10 lines × 3) (apps/omlx-mac/Sources/AppView/Screens/AppearanceScreen.swift)
- Duplicated block (10 lines × 4) (apps/omlx-mac/Sources/AppView/Screens/NetworkScreen.swift)
- Duplicated block (10 lines × 9) (apps/omlx-mac/Sources/AppView/Screens/ModelSettingsScreen.swift)
- Duplicated block (10–11 lines × 2) (apps/omlx-mac/Sources/AppView/Screens/IntegrationsScreen.swift)
- Duplicated block (10–12 lines × 3) (omlx/cluster/incidents.py)
- Duplicated block (10–15 lines × 3) (apps/omlx-mac/Sources/AppView/Screens/AppearanceScreen.swift)
- Duplicated block (11 lines × 5) (apps/omlx-mac/Sources/AppView/Screens/AppearanceScreen.swift)
- Duplicated block (11–14 lines × 2) (apps/omlx-mac/Sources/AppView/Screens/DownloadsScreen.swift)
- Duplicated block (11–16 lines × 4) (apps/omlx-mac/Sources/AppView/Screens/AppearanceScreen.swift)
- Duplicated block (12 lines × 2) (apps/omlx-mac/Sources/AppView/AppView.swift)
- Duplicated block (12 lines × 2) (apps/omlx-mac/Sources/AppView/Screens/ModelsScreen.swift)
- Duplicated block (12 lines × 2) (apps/omlx-mac/Sources/AppView/Screens/ModelsScreen.swift)
- Duplicated block (12 lines × 2) (apps/omlx-mac/Sources/AppView/Screens/QuantizationScreen.swift)
- …and 129 more

## New (428)

- Attention.__call__ (cognitive 16) (apps/omlx-cuda-helper/omlx_cuda_helper/families/qwen4_exp/attention.py)
- Change-coupling hub: dashboard.js → ModelSettingsScreen.swift, PerformanceScreen.swift, ModelSettingsScreenVM.swift, ModelsDTO.swift, accuracy_benchmark.py, routes.py, config.py, settings.py (apps/omlx-web/omlx_web/static/js/dashboard.js)
- ClassTooLong: IntegrationsScreenVM (apps/omlx-mac/Sources/AppView/ViewModels/IntegrationsScreenVM.swift)
- Config.validate (cognitive 17) (apps/omlx-cuda-helper/omlx_cuda_helper/families/deepseek_v41/config.py)
- ControlServer._handle (cognitive 16) (apps/omlx-cuda-helper/omlx_cuda_helper/control.py)
- ControlServer._handle (cyclomatic 16) (apps/omlx-cuda-helper/omlx_cuda_helper/control.py)
- Duplicated block (10 lines × 2) (omlx/admin/routes.py)
- Duplicated block (10 lines × 2) (omlx/patches/inkling2/model.py)
- Duplicated block (10–18 lines × 2) (apps/omlx-mac/Sources/Updater/AppUpdater.swift)
- Duplicated block (11 lines × 2) (apps/omlx-cuda-helper/omlx_cuda_helper/wire.py)
- Duplicated block (11 lines × 2) (omlx/engine/batched.py)
- Duplicated block (11 lines × 2) (omlx/patches/inkling2/loader.py)
- Duplicated block (11 lines × 2) (omlx/patches/inkling2/oq_virtual.py)
- Duplicated block (12 lines × 2) (apps/omlx-cuda-helper/omlx_cuda_helper/families/deepseek_v41/ngram.py)
- Duplicated block (12 lines × 2) (apps/omlx-cuda-helper/omlx_cuda_helper/worker.py)
- Duplicated block (13 lines × 2) (apps/omlx-cuda-helper/omlx_cuda_helper/families/glm5_next/model.py)
- Duplicated block (13 lines × 2) (apps/omlx-cuda-helper/omlx_cuda_helper/families/qwen3_5/stage.py)
- Duplicated block (13 lines × 2) (apps/omlx-cuda-helper/omlx_cuda_helper/wire.py)
- Duplicated block (13 lines × 2) (omlx/patches/inkling2/dspark.py)
- Duplicated block (13 lines × 3) (omlx/memory_monitor.py)
- …and 408 more

## Changes since last survey

- 118 commits — 56 feature/other, 62 fixes

## By area

- apps/omlx-web — 22 commits
- apps/omlx-mac — 12 commits
- omlx/patches — 12 commits
- apps/omlx-cuda-helper — 11 commits
- omlx/admin — 11 commits
- omlx/api — 8 commits
- omlx/engine — 7 commits
- (root) — 3 commits
- omlx/affine4.py — 3 commits
- .github/workflows — 2 commits
- Formula/omlx.rb — 2 commits
- omlx/adapter — 2 commits
- omlx/cluster — 2 commits
- omlx/engine_core.py — 2 commits
- omlx/integrations — 2 commits
- omlx/prefill_helper — 2 commits
- omlx/scheduler.py — 2 commits
- docs/Splash.md — 1 commit
- docs/TESTING.md — 1 commit
- omlx/_version.py — 1 commit

## Notable commits

- fix: fix(adapter): keep model text from forging harmony log lines (#5125)
- fix: fix(admin): check drafts when a profile is applied (#5105)
- fix: fix(admin): name the refused field when a settings save is rejected (#4513)
- fix: fix(admin): report MTP as unused when an external engine serves the model (#5168)
- fix: fix(admin): show recent decode speed per request, not the lifetime average (#5075)
- fix: fix(admin): stop a parked draft from blocking settings and profile saves (#4401)
- fix: fix(affine): keep attention sinks on the fused bf16 prefill path (#5088)
- fix: fix(api): keep the usual key order in tool schemas (#5161)
- fix: fix(api): surface a coded error when a tool_calls finish has no calls (#5139)
- fix: fix(bench): grow a short saved Splash context for the run (#4938)
- fix: fix(benchmark): batch tests wider than max concurrent requests report inflated tg TPS (#5312)
- fix: fix(benchmark): include gen_tps in the pp result log line (#5126)
- fix: fix(benchmark): pin Prefill First for benchmark runs
- fix: fix(brew): install the cluster extra so the Cluster tab finds NVIDIA prefill helpers
- fix: fix(cluster): name the link, not the head counts, when the link gate blocks TP (#3024)
- fix: fix(cluster): pass DeepSeek V4.1 tool arguments to mlx-lm as objects
- fix: fix(codex): skip the reverse DNS lookup in the instructions probe server
- fix: fix(cuda-helper): keep the bf16 roundings of Flash-Next kernels on sm_90 and newer
- fix: fix(cuda-helper): leave the OS 16 GiB when an integrated GPU plans its layers
- fix: fix(dflash): restore drafter context on exact prefix hits
- …and 98 more

## Architecture

- Unchanged — 0 containers · 1 contexts · 0 edges
