v2.5.0

2.5.0 (2026-10-22)

Features

• api: model-aware K/V support and runtime status endpoints (b6dff8e)
• binary: build llama.cpp from source with detected toolchains (b6dff8e)
• binary: journaled runtime promotion with rollback and recovery (b6dff8e)
• inference: capability-aware mixed K/V classification (b6dff8e)
• ui: llama.cpp runtime panel for builds, updates and rollback (b6dff8e)

Bug Fixes

• binary: harden build supervision and runtime swap ownership (b6dff8e)
• binary: stop scanning migration receipts byte by byte (a297b9a)

v2.4.1

2.4.1 (2026-10-10)

Bug Fixes

• certs: drop verifier result from test assertion messages (aa936df)

v2.4.0

2.4.0 (2026-10-09)

Features

• models: download Rapid-MLX models from cards and reuse the Hugging Face cache (357bf4c)
• nav: show engine identity and distinguish model repositories from served aliases (357bf4c)
• ui: integrate Rapid-MLX telemetry into shared dashboard cards and active requests (357bf4c)
• ui: show model quantization and memory estimates on preset cards (357bf4c)
• wizard: add Rapid-MLX download steps and context-length controls (357bf4c)

Bug Fixes

• api: harden Rapid-MLX model lookups and download job identifiers (357bf4c)
• ci: give UI builds and test jobs independent 90-minute timeout budgets (a682e85)
• ci: reuse exact-commit UI build artifacts across isolated test shards and reruns (a682e85)
• ci: save PR-scoped Rust caches while retaining main dependency fallback (a682e85)
• deps: update rust crate zip to v9 (#452) (ed88186)
• gpu: distinguish unavailable CPU and compact GPU samples from genuine zero load (357bf4c)
• models: reject incomplete snapshots without blocking valid caches for unrelated partial downloads (357bf4c)
• models: retain download state across surfaces and recover transient errors and missing jobs (357bf4c)
• settings: keep TLS certificate controls clickable and stable (357bf4c)
• spawn: filter routine backend health and status polling logs (357bf4c)
• spawn: stop inference backends safely on SIGINT, SIGTERM, and tray Quit (357bf4c)
• spawn: use bf16 KV for hybrid Rapid-MLX models and validate tool parser settings (357bf4c)
• spawn: use the selected managed Rapid runtime for hybrid model checks (357bf4c)
• ui: preserve active-request selection and hide retained telemetry after failed polls (357bf4c)
• ui: respect light themes and reduced motion in Rapid-MLX controls (357bf4c)
• wizard: preserve Rapid Auto context and saved template reasoning effort (357bf4c)
• wizard: use safe Rapid-MLX concurrency defaults and consistent model controls (357bf4c)

v2.3.1

2.3.1 (2026-10-07)

Bug Fixes

• wizard: keep Guided vision projector selection and VRAM size (c3db4e5)
• wizard: sync hardware state when entering the Launch step (c3db4e5)

v2.3.0

2.3.0 (2026-10-07)

Features

• dashboard: live throughput and prefill progress for llama.cpp (d285a42)
• dashboard: llama.cpp efficiency telemetry and runtime details (d285a42)
• dashboard: Strata-style state strip and unified metric cards (d285a42)
• hf: refresh Discover pills for the current model landscape (d285a42)
• wizard: raise default max tokens and reasoning budget (d285a42)

Bug Fixes

• api: bound upstream waits without limiting healthy streams (d285a42)
• attach: actionable error when a LAN runtime is unreachable (d285a42)
• attach: log the full error chain when a runtime health check fails (d285a42)
• chat: persist context fingerprints so stale usage is not reused (d285a42)
• ctx: correct context math; cockpit G-first order and fit (d285a42)
• ctx: use occupancy formula in the send-path overflow guard (d285a42)
• dashboard: 5-min sparklines, PP freeze, honest GPU load (d285a42)
• dashboard: drop inactive phase to baseline, freeze on idle (d285a42)
• dashboard: match Strata speed sparkline semantics exactly (d285a42)
• dashboard: quantize PP/TG rate reports to counter chunks (d285a42)
• dashboard: stable state classification and true PP rate (d285a42)
• dashboard: strata card corrections for Apple Silicon (d285a42)
• dashboard: Strata-style speed card and EMA-smoothed rates (d285a42)
• dashboard: stronger EMA smoothing on the prefill rate (d285a42)
• dashboard: throttle speed samples to 1/s for a true 5-min window (d285a42)
• gpu: parse Apple ioreg utilization from PerformanceStatistics only (d285a42)
• gpu: real Apple GPU utilization from IOAccelerator (d285a42)
• metrics: true CPU utilization; abort-proof chat timeouts (d285a42)
• ui: cockpit chips never collapse; advanced panel fits (d285a42)
• wizard: drawer access for sampling, template card, sticky bar (d285a42)
• wizard: keep Guided and Pro spawn wizard state in sync (d285a42)
• wizard: restore mmproj discovery in Guided; hide review block (d285a42)
• wizard: stop duplicating guided decisions (d285a42)

v2.2.1

2.2.1 (2026-10-06)

Bug Fixes

• ci: handle escaped newlines in macOS signing secrets (7545392)

v2.2.0

2.2.0 (2026-10-05)

Features

• ci: sign and notarize the macOS build (d852d96)

Bug Fixes

• deps: update rust dependencies to v0.26.6 (#432) (c258a49)

v2.1.2

2.1.2 (2026-09-19)

Bug Fixes

• deps: update rust crate dirs to v7 (#420) (7ecfc21)
• deps: update rust crate tray-icon to 0.25 (#415) (c2f755d)
• deps: update rust crate wry to 0.57 (#416) (979051b)
• wizard: isolate advisor metadata by model (ac27757)

v2.1.1

2.1.1 (2026-09-03)

Bug Fixes

• release: align release workflow with legacy-asset drop policy (#382) (509d66d)

v2.1.0

2.1.0 (2026-09-03)

Features

• calibration: implement WddmTotalDeviceDelta nvidia-smi sampler (44226a5)
• presets: add bounded llama-fit-params probe (44226a5)
• presets: add evidence-details drawer action for bundle launch memory (44226a5)
• presets: add two-sided n_cpu_moe placement search (44226a5)
• presets: add typed launch bundles and reasoning flags (44226a5)
• presets: manage explicit model artifact bundles (44226a5)
• presets: render bundle presets as a single compact launch card (44226a5)
• ui: add preset bundle configure drawer (44226a5)
• vram: back launch estimates with fit probe (44226a5)
• vram: launch-evidence vocabulary, store, and resolve wiring (44226a5)
• vram: resolve explicit preset fit intents (44226a5)
• wizard: expose canonical llama runtime controls (44226a5)

Bug Fixes

• api: don't commit preset mutations to memory when save fails (44226a5)
• api: run mixed K/V hard-gate check in Doctor independent of tool_enabled (44226a5)
• api: surface non-blocking K/V policy signal on vram-estimate (44226a5)
• brand: center the social-preview export correctly (44226a5)
• calibration: bound telemetry capture with a timeout (44226a5)
• calibration: enforce K/V launch policy gate at every calibration launch site (44226a5)
• calibration: fix flaky apply-validation fixture under parallel test load (44226a5)
• calibration: repeated-cycle agreement for metal_sampler (44226a5)
• capabilities: stop re-deriving help_hash from serve_flags at lookup (44226a5)
• inference: remove fail-closed gate on reasoning-preserve flag (44226a5)
• presets: close two Greptile-flagged bundle-launch races (44226a5)
• presets: enforce canonical llama launch validation (44226a5)
• presets: fix reachable resolve panic and extract shared cycle-agreement helper (44226a5)
• presets: remove dead fallback and close bundle.rs test gaps (44226a5)
• vram: enforce exact preset intent provenance (44226a5)
• vram: preserve context for low-vram intents (44226a5)
• vram: round layer-split fallback, handle nvidia-smi N/A, fix wired-limit tier gap (44226a5)