RPC: populate use_count to enable fusion inside backends (#27142)

* RPC: populate use_count to enable fusion inside backends

* bump minor version

.devops/openvino.Dockerfile

@@ -1,18 +1,18 @@
-ARG OPENVINO_VERSION_MAJOR=2026.2.1
-ARG OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3
+ARG OPENVINO_VERSION_MAJOR=2026.3
+ARG OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c
 ARG UBUNTU_VERSION=24.04
 
 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.36.3
-ARG IGC_VERSION_FULL=2_2.36.3+21719
-ARG COMPUTE_RUNTIME_VERSION=26.22.38646.4
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.22.38646.4-0
+ARG IGC_VERSION=v2.38.2
+ARG IGC_VERSION_FULL=2_2.38.2+22051
+ARG COMPUTE_RUNTIME_VERSION=26.27.39122.11
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.27.39122.11-0
 ARG IGDGMM_VERSION=22.10.0
 
 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
-ARG NPU_DRIVER_VERSION=v1.33.0
-ARG NPU_DRIVER_FULL=v1.33.0.20260529-26625960453
-ARG LIBZE1_VERSION=1.27.0-1~24.04~ppa2
+ARG NPU_DRIVER_VERSION=v1.35.0
+ARG NPU_DRIVER_FULL=v1.35.0.20260722-29947505341
+ARG LIBZE1_VERSION=1.28.2-1~24.04~ppa1
 
 # Optional proxy build arguments
 ARG http_proxy=
@@ -170,7 +170,7 @@ RUN --mount=type=cache,target=/var/cache/intel-npu,sharing=locked \
     fi; \
     DEB=/var/cache/intel-npu/libze1_${LIBZE1_VERSION}_amd64.deb; \
     if [ ! -f "$DEB" ]; then \
-        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260324T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
+        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260606T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
     fi; \
     mkdir /tmp/npu/ && cd /tmp/npu/ && tar -xf "$TGZ" && cp "$DEB" .; \
     apt-get update; \

.github/actions/windows-setup-cuda/action.yml

@@ -6,8 +6,7 @@ inputs:
     required: true
   cuda_arch:
     description: "CUDA target architecture"
-    required: false
-    default: "x64"
+    required: true
 
 runs:
   using: "composite"

.github/workflows/build-cache.yml

@@ -40,9 +40,9 @@ jobs:
     runs-on: ubuntu-24.04
 
     env:
-      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone
@@ -69,8 +69,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone

.github/workflows/build-cuda-windows.yml

@@ -22,6 +22,7 @@ env:
 
 jobs:
   cuda:
+    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
     runs-on: windows-2022
 
     permissions:
@@ -29,7 +30,16 @@ jobs:
 
     strategy:
       matrix:
-        cuda: ['12.4', '13.3']
+        include:
+          - cuda: '12.4'
+            arch: x64
+            defines: '-DGGML_CUDA_CUB_3DOT2=ON'
+          - cuda: '13.3'
+            arch: x64
+            defines: ''
+          - cuda: '13.4'
+            arch: arm64
+            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'
 
     steps:
       - name: Clone
@@ -39,12 +49,13 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
 
       - name: Install Cuda Toolkit
         uses: ./.github/actions/windows-setup-cuda
         with:
           cuda_version: ${{ matrix.cuda }}
+          cuda_arch: ${{ matrix.arch }}
 
       - name: Install Ninja
         id: install_ninja
@@ -54,26 +65,21 @@ jobs:
       - name: Build
         id: cmake_build
         shell: cmd
-        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
           cmake -S . -B build -G "Ninja Multi-Config" ^
-            -DLLAMA_BUILD_SERVER=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_NATIVE=OFF ^
             -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_CPU=OFF ^
             -DGGML_CUDA=ON ^
-            -DGGML_RPC=ON ^
-            -DGGML_CUDA_CUB_3DOT2=ON
+            -DLLAMA_BUILD_BORINGSSL=ON ${{ matrix.defines }}
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
-          cmake --build build --config Release
+          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda
 
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
         with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
 
   hip:
     runs-on: windows-2022

.github/workflows/build-openvino.yml

@@ -39,8 +39,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone
@@ -81,7 +81,7 @@ jobs:
         # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           cd ${{ github.workspace }}
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 2000
 
       - name: Test (GPU)
         id: cmake_test_gpu
@@ -89,15 +89,15 @@ jobs:
         run: |
           cd ${{ github.workspace }}
           export GGML_OPENVINO_DEVICE=GPU
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 3000
 
   openvino-windows-2022:
     runs-on: windows-2022
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone
@@ -166,4 +166,4 @@ jobs:
           call "%OPENVINO_ROOT%\setupvars.bat"
 
           cd build
-          ctest --test-dir ReleaseOV -L main -E "test-llama-archs" -C Release --verbose --timeout 3000
+          ctest --test-dir ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" -C Release --verbose --timeout 3000

.github/workflows/build-self-hosted.yml

@@ -288,8 +288,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone

.github/workflows/release.yml

@@ -446,8 +446,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Set OpenVINO version output
@@ -562,8 +562,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Set OpenVINO version output

README.md

@@ -7,10 +7,11 @@
 <b>LLM inference in C/C++</b>
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp?filter=v*)](https://github.com/ggml-org/llama.cpp/releases?q=tag:v0)
+[![Nightly](https://img.shields.io/github/v/release/ggml-org/llama.cpp?label=nightly)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
-[![Docker](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
-[![Winget](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)
+[![Docker](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/docker.yml?label=Docker)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
+[![Winget](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/winget.yml?label=Winget)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)
 
 [manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
 

ci/run.sh

@@ -190,7 +190,7 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"
 
     # TODO: fix and re-enable the `test-llama-archs` test below
-    CTEST_EXTRA="-E test-llama-archs"
+    CTEST_EXTRA="-E test-llama-archs|test-recurrent-state-rollback-nemotron-h"
 fi
 
 ## helpers

common/common.cpp

@@ -1750,6 +1750,18 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const commo
     return tpp;
 }
 
+namespace {
+
+bool can_share_threadpool(const ggml_threadpool_params & tpp1, const ggml_threadpool_params & tpp2) {
+    // n_threads does not matter -> we'll use what's larger
+    ggml_threadpool_params tpp_comparison = tpp1;
+    tpp_comparison.n_threads = tpp2.n_threads;
+
+    return ggml_threadpool_params_match(&tpp_comparison, &tpp2);
+}
+
+}  // namespace
+
 common_threadpools::~common_threadpools() {
     if (!free_fn) {
         return;
@@ -1778,7 +1790,9 @@ void common_threadpools::init(llama_context * ctx, const common_params & params)
     struct ggml_threadpool_params tpp =
             ggml_threadpool_params_from_cpu_params(params.cpuparams);
 
-    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
+    if (can_share_threadpool(tpp, tpp_batch)) {
+        tpp.n_threads = std::max(tpp.n_threads, tpp_batch.n_threads);
+    } else {
         threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
         if (!threadpool_batch) {
             COM_WRN("batch threadpool create failed : n_threads %d\n", tpp_batch.n_threads);

docs/backend/OPENVINO.md

@@ -237,8 +237,8 @@ chmod +x ubuntu-llamacpp-ov-install.sh
 # ============================================
 set -euo pipefail
 
-OPENVINO_VERSION_MAJOR="2026.2.1"
-OPENVINO_VERSION_FULL="2026.2.1.21919.ede283a88e3"
+OPENVINO_VERSION_MAJOR="2026.3"
+OPENVINO_VERSION_FULL="2026.3.0.22451.bd8d6542e3c"
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -334,7 +334,7 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```
 
 > [!NOTE]
-> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
 
 </details>
 
@@ -364,8 +364,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================
 
-set "OPENVINO_VERSION_MAJOR=2026.2.1"
-set "OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3"
+set "OPENVINO_VERSION_MAJOR=2026.3"
+set "OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c"
 
 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -547,7 +547,7 @@ endlocal
 ```
 
 > [!NOTE]
-> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
 
 </details>
 

ggml/include/ggml-rpc.h

@@ -7,7 +7,7 @@ extern "C" {
 #endif
 
 #define RPC_PROTO_MAJOR_VERSION    5
-#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_MINOR_VERSION    1
 #define RPC_PROTO_PATCH_VERSION    0
 
 #ifdef  __cplusplus

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -47,7 +47,7 @@ struct rpc_tensor {
     uint64_t data;
     char name[GGML_MAX_NAME];
 
-    char padding[4];
+    int32_t use_count;
 };
 
 static_assert(sizeof(rpc_tensor) % 8 == 0, "rpc_tensor size must be multiple of 8");
@@ -447,7 +447,7 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
 
     // Avoid sending uninitialized data over the wire
     memset(result.name, 0, sizeof(result.name));
-    memset(result.padding, 0, sizeof(result.padding));
+    result.use_count = 0;
 
     snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
     return result;
@@ -675,7 +675,7 @@ static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
     // this is no-op because we don't have any async operations
 }
 
-static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
+static void add_tensor(ggml_tensor * tensor, const ggml_cgraph * cgraph, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
     if (tensor == nullptr) {
         return;
     }
@@ -684,10 +684,15 @@ static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors,
     }
     visited.insert(tensor);
     for (int i = 0; i < GGML_MAX_SRC; i++) {
-        add_tensor(tensor->src[i], tensors, visited);
+        add_tensor(tensor->src[i], cgraph, tensors, visited);
     }
-    add_tensor(tensor->view_src, tensors, visited);
-    tensors.push_back(serialize_tensor(tensor));
+    add_tensor(tensor->view_src, cgraph, tensors, visited);
+    rpc_tensor result = serialize_tensor(tensor);
+    const size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, tensor);
+    if (hash_pos != GGML_HASHSET_FULL && ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos)) {
+        result.use_count = cgraph->use_counts[hash_pos];
+    }
+    tensors.push_back(result);
 }
 
 static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::vector<uint8_t> & output) {
@@ -695,7 +700,7 @@ static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::ve
     std::vector<rpc_tensor> tensors;
     std::unordered_set<ggml_tensor*> visited;
     for (uint32_t i = 0; i < n_nodes; i++) {
-        add_tensor(cgraph->nodes[i], tensors, visited);
+        add_tensor(cgraph->nodes[i], cgraph, tensors, visited);
     }
     // serialization format:
     // | device (4 bytes) | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
@@ -1451,6 +1456,10 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input) {
             GGML_LOG_ERROR("[%s] failed to create graph node %d (id=%" PRId64 ")\n", __func__, i, id);
             return false;
         }
+        if (graph->nodes[i] != nullptr) {
+            const size_t hash_pos = ggml_hash_insert(&graph->visited_hash_set, graph->nodes[i]);
+            graph->use_counts[hash_pos] = tensor_ptrs.at(id)->use_count;
+        }
     }
     ggml_status status = ggml_backend_graph_compute(backends[device], graph);
     GGML_ASSERT(status == GGML_STATUS_SUCCESS && "Unsuccessful graph computations are not supported with RPC");

ggml/src/ggml-sycl/fwht.cpp

@@ -0,0 +1,119 @@
+#include "fwht.hpp"
+
+#include <cmath>
+
+template <int N>
+static void fwht_kernel(const float * __restrict__ src, float * __restrict__ dst, const int64_t n_rows,
+                        const float scale, const sycl::nd_item<2> & item) {
+    const sycl::sub_group sg = item.get_sub_group();
+
+    const int64_t r = item.get_global_id(0);
+    if (r >= n_rows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    constexpr int el_w = N / WARP_SIZE;
+    static_assert(el_w >= 1 && N % WARP_SIZE == 0, "row must be a whole number of sub-group widths");
+
+    float     reg[el_w];
+    const int lane = sg.get_local_linear_id();
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        reg[i] = src[i * WARP_SIZE + lane] * scale;
+    }
+
+    // Butterflies inside the sub-group. The partner of a lane with bit h clear is the
+    // lower index of the pair, so it takes the sum and the upper takes lower - upper.
+#pragma unroll
+    for (int h = 1; h < WARP_SIZE; h *= 2) {
+#pragma unroll
+        for (int j = 0; j < el_w; ++j) {
+            const float val  = reg[j];
+            const float val2 = dpct::permute_sub_group_by_xor(sg, val, h, WARP_SIZE);
+
+            reg[j] = (lane & h) == 0 ? val + val2 : val2 - val;
+        }
+    }
+
+    // Butterflies across registers: h is a multiple of WARP_SIZE, so the partner of
+    // element i*WARP_SIZE + lane lives in reg[i + h/WARP_SIZE] on the same lane.
+#pragma unroll
+    for (int h = WARP_SIZE; h < N; h *= 2) {
+        const int step = h / WARP_SIZE;
+#pragma unroll
+        for (int j = 0; j < el_w; j += 2 * step) {
+#pragma unroll
+            for (int k = 0; k < step; ++k) {
+                const float x = reg[j + k];
+                const float y = reg[j + k + step];
+
+                reg[j + k]        = x + y;
+                reg[j + k + step] = x - y;
+            }
+        }
+    }
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        dst[i * WARP_SIZE + lane] = reg[i];
+    }
+}
+
+template <int N>
+static void launch_fwht(const float * src, float * dst, const int64_t n_rows, const float scale,
+                        dpct::queue_ptr stream) {
+    constexpr int rows_per_block = 4;
+
+    const int64_t num_blocks = (n_rows + rows_per_block - 1) / rows_per_block;
+
+    // dim 1 is the fastest-varying, so a sub-group is exactly one row's WARP_SIZE lanes.
+    const sycl::range<2> global(num_blocks * rows_per_block, WARP_SIZE);
+    const sycl::range<2> local(rows_per_block, WARP_SIZE);
+
+    stream->parallel_for(sycl::nd_range<2>(global, local),
+                         [=](sycl::nd_item<2> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             fwht_kernel<N>(src, dst, n_rows, scale, item);
+                         });
+}
+
+bool ggml_sycl_op_fwht(ggml_backend_sycl_context & ctx, const ggml_tensor * src, ggml_tensor * dst) {
+    if (src->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+    if (!ggml_are_same_shape(src, dst)) {
+        return false;
+    }
+    if (!ggml_is_contiguous(src) || !ggml_is_contiguous(dst)) {
+        return false;
+    }
+
+    const int     n    = (int) src->ne[0];
+    const int64_t rows = ggml_nrows(src);
+
+    const float *   src_d  = (const float *) src->data;
+    float *         dst_d  = (float *) dst->data;
+    dpct::queue_ptr stream = ctx.stream();
+
+    const float scale = 1.0f / std::sqrt((float) n);
+
+    switch (n) {
+        case 64:
+            launch_fwht<64>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 128:
+            launch_fwht<128>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 256:
+            launch_fwht<256>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 512:
+            launch_fwht<512>(src_d, dst_d, rows, scale, stream);
+            return true;
+        default:
+            return false;
+    }
+}

ggml/src/ggml-sycl/fwht.hpp

@@ -0,0 +1,12 @@
+#ifndef GGML_SYCL_FWHT_HPP
+#define GGML_SYCL_FWHT_HPP
+
+#include "common.hpp"
+
+// Fast Walsh-Hadamard transform, the fast path for a MUL_MAT whose src0 ggml has
+// tagged GGML_HINT_SRC0_IS_HADAMARD. src0 is not read at all. Returns false if the
+// shape is not one this can serve, in which case the caller must fall through to the
+// ordinary mat-mul dispatch.
+bool ggml_sycl_op_fwht(ggml_backend_sycl_context & ctx, const ggml_tensor * src, ggml_tensor * dst);
+
+#endif  // GGML_SYCL_FWHT_HPP

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -58,6 +58,7 @@
 #include "ggml-sycl/backend.hpp"
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/element_wise.hpp"
+#include "ggml-sycl/fwht.hpp"
 #include "ggml-sycl/gemm.hpp"
 #include "ggml-sycl/getrows.hpp"
 #include "ggml-sycl/norm.hpp"
@@ -4473,6 +4474,18 @@ static bool can_use_mul_mat_vec_q(const ggml_tensor * src0, const ggml_tensor *
 
 static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    // Handle HADAMARAD hint given from further up the pipeline and pass it to the correct
+    // kernel.
+    //
+    // The op check is not redundant: this backend also routes MUL_MAT_ID through here with a
+    // stack copy of dst, which carries MUL_MAT_ID's own op_params. ggml_mul_mat_set_hint()
+    // asserts GGML_OP_MUL_MAT for the same reason.
+    if (dst->op == GGML_OP_MUL_MAT && ggml_get_op_params_i32(dst, 1) == GGML_HINT_SRC0_IS_HADAMARD &&
+        ggml_sycl_op_fwht(ctx, src1, dst)) {
+        return;
+    }
+
     const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
     int64_t min_compute_capability = INT_MAX;
 

src/llama-model-loader.cpp

@@ -1395,6 +1395,11 @@ void llama_model_loader::get_mapping_range(size_t * first, size_t * last, void *
     }
 }
 
+void llama_model_loader::unmap_weight(const llama_tensor_weight & w) const {
+    if (!use_mmap) { return; }
+    mappings.at(w.idx)->unmap_fragment(w.offs, w.offs + ggml_nbytes(w.tensor));
+}
+
 void llama_model_loader::load_data_for(struct ggml_tensor * cur) const {
     const auto & w = require_weight(ggml_get_name(cur));
 

src/llama-model-loader.h

@@ -194,6 +194,9 @@ struct llama_model_loader {
 
     void get_mapping_range(size_t * first, size_t * last, void ** addr, int idx, ggml_context * ctx) const;
 
+    // release a weight's mmap pages
+    void unmap_weight(const llama_tensor_weight & w) const;
+
     // for backwards compatibility, does not support ggml-backend
     void load_data_for(struct ggml_tensor * cur) const;
 

src/llama-quant.cpp

@@ -1270,7 +1270,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             total_size_org += tensor_size;
             total_size_new += new_size;
 
-            // update the gguf meta data as we go
+            // update the gguf metadata as we go
             gguf_set_tensor_type(ctx_outs[cur_split].get(), metadata[i].name.c_str(), new_type);
             GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), metadata[i].name.c_str())) == new_size);
             gguf_set_tensor_data(ctx_outs[cur_split].get(), metadata[i].name.c_str(), new_data);
@@ -1278,6 +1278,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             // write tensor data + padding
             fout.write((const char *) new_data, new_size);
             zeros(fout, GGML_PAD(new_size, align) - new_size);
+
+            // unmap the tensor to free memory
+            if (ml.use_mmap) { ml.unmap_weight(weight); }
+
         } // no --dry-run
     } // main loop
 

src/unicode.cpp

@@ -1241,7 +1241,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
         { unicode_cpt_flags::LETTER,      "\x41-\x5A\x61-\x7A" }, // A-Za-z
         { unicode_cpt_flags::PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
         { unicode_cpt_flags::ACCENT_MARK, "" }, // no sub-128 codepoints
-        { unicode_cpt_flags::SYMBOL,      "\\\x24\\\x2B\x3C-\x3E\x5E\x60\\\x7C" }, // $+<=>^`|
+        { unicode_cpt_flags::SYMBOL,      "\\\x24\\\x2B\x3C-\x3E\x5E\x60\\\x7C\\\x7E" }, // $+<=>^`|~
     };
 
     // compute collapsed codepoints only if needed by at least one regex

tests/CMakeLists.txt

@@ -116,6 +116,8 @@ function(llama_build_and_test source)
     set_property(TEST ${TEST_TARGET} PROPERTY LABELS ${LLAMA_TEST_LABEL})
 endfunction()
 
+llama_build_and_test(test-unicode.cpp)
+
 # build test-tokenizer-0 target once and add many tests
 llama_build(test-tokenizer-0.cpp)
 
@@ -310,6 +312,9 @@ llama_build_and_test(test-mtmd-c-api.c)
 target_link_libraries(${LLAMA_TEST_NAME} PRIVATE mtmd)
 unset(LLAMA_TEST_NAME)
 
+llama_build_and_test(test-mtmd-impl.cpp)
+target_link_libraries(test-mtmd-impl PRIVATE mtmd)
+
 # GGUF model data fetcher library for tests that need real model metadata
 # Only compile when cpp-httplib has SSL support (CPPHTTPLIB_OPENSSL_SUPPORT)
 if (TARGET cpp-httplib)

tests/test-mtmd-impl.cpp

@@ -0,0 +1,88 @@
+#include "testing.h"
+
+#include "mtmd-image.h"
+
+#include <iostream>
+#include <string>
+#include <utility>
+#include <vector>
+
+// this test file contains:
+// 1. test cases for mtmd helpers
+// 2. test cases for internal mtmd components
+// internal headers can be included here
+
+struct test_registry {
+    using fn_t = void (*)(testing &);
+
+    struct entry {
+        std::string name;
+        fn_t fn;
+    };
+
+    static std::vector<entry> & all() {
+        static std::vector<entry> entries;
+        return entries;
+    }
+
+    test_registry(const char * name, fn_t fn) {
+        all().push_back({ name, fn });
+    }
+};
+
+#define MAKE_TEST(name)                                               \
+    static void name(testing & t);                                    \
+    static const test_registry test_registry_ ## name(#name, &name);  \
+    static void name(testing & t)
+
+
+//
+// mtmd_image
+//
+
+MAKE_TEST(test_image_preprocessor_lfm2) {
+    clip_hparams hparams;
+    hparams.patch_size = 16;
+    hparams.n_merge = 2;
+    hparams.set_limit_image_tokens(64, 256);
+
+    // { image size, expected tiling }
+    const std::vector<std::pair<clip_image_size, bool>> cases = {
+        { {  704, 704 }, false },
+        // 720 / (patch_size * n_merge) is exactly 22.5, so this only matches HF
+        // if round_by_factor rounds half to even (22) instead of away from zero (23)
+        { {  720, 720 }, false },
+        { {  736, 736 }, true  },
+        { { 1024, 977 }, true  },
+        { { 1056, 384 }, false },
+    };
+
+    for (const auto & [size, expected] : cases) {
+        const bool actual = mtmd_image_preprocessor_lfm2::should_tile(hparams, size);
+
+        t.assert_equal(
+            "tiling for " + std::to_string(size.width) + "x" + std::to_string(size.height),
+            std::string(expected ? "tiled" : "single"),
+            std::string(actual   ? "tiled" : "single"));
+    }
+}
+
+//
+// main
+//
+
+int main(int argc, char ** argv) {
+    testing t(std::cout);
+    t.verbose = true;
+
+    // usage: test-mtmd-impl [filter_regex]
+    for (int i = 1; i < argc; i++) {
+        t.set_filter(argv[i]);
+    }
+
+    for (const auto & e : test_registry::all()) {
+        t.test(e.name, e.fn);
+    }
+
+    return t.summary();
+}

tests/test-unicode.cpp

@@ -0,0 +1,24 @@
+#include "../src/unicode.h"
+
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main() {
+    const std::vector<std::string> regex_exprs = {
+        "[~][A-Za-z]+| ?[\\p{S}]+|\\s+",
+    };
+    const std::vector<std::string> expected = { " ~", "foo" };
+    const auto actual = unicode_regex_split(" ~foo", regex_exprs, false);
+
+    if (actual != expected) {
+        fprintf(stderr, "unexpected split:");
+        for (const auto & piece : actual) {
+            fprintf(stderr, " [%s]", piece.c_str());
+        }
+        fprintf(stderr, "\n");
+        return 1;
+    }
+
+    return 0;
+}

tools/mtmd/CMakeLists.txt

@@ -90,6 +90,9 @@ if (BUILD_SHARED_LIBS)
     set_target_properties     (mtmd PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_compile_definitions(mtmd PRIVATE LLAMA_BUILD)
     target_compile_definitions(mtmd PUBLIC  LLAMA_SHARED)
+
+    # export all symbols so that internal components can be tested by test-mtmd-impl
+    set_target_properties     (mtmd PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS ON)
 endif()
 
 set(MTMD_PUBLIC_HEADERS

tools/mtmd/clip-impl.h

@@ -858,6 +858,9 @@ static std::ifstream open_ifstream_binary(const std::string & fname) {
 }
 #endif
 
+// in test-mtmd-impl, we include woth common.h and this file, and these functions are duplicated
+// this is a quick fix to avoid compilation errors
+#ifndef DIRECTORY_SEPARATOR
 static std::string string_format(const char * fmt, ...) {
     va_list ap;
     va_list ap2;
@@ -915,6 +918,7 @@ inline bool string_ends_with(std::string_view str, std::string_view suffix) {
     return str.size() >= suffix.size() &&
            str.compare(str.size() - suffix.size(), suffix.size(), suffix) == 0;
 }
+#endif
 
 //
 // gguf utils

tools/mtmd/mtmd-image.cpp

@@ -1013,14 +1013,31 @@ mtmd_image_preproc_out mtmd_image_preprocessor_lfm2::preprocess(const clip_image
     return output;
 }
 
+bool mtmd_image_preprocessor_lfm2::should_tile(
+        const clip_hparams & hparams,
+        const clip_image_size & original_size) {
+    const int align_size = hparams.patch_size * hparams.n_merge;
+
+    const auto round_by_factor = [align_size](float x) {
+        // see https://github.com/ggml-org/llama.cpp/pull/27057#discussion_r3796264887
+        return static_cast<int>(std::nearbyint(static_cast<double>(x) / align_size)) * align_size;
+    };
+
+    const int h_bar = std::max(hparams.patch_size, round_by_factor(original_size.height));
+    const int w_bar = std::max(hparams.patch_size, round_by_factor(original_size.width));
+
+    return static_cast<double>(h_bar) * static_cast<double>(w_bar) >
+           static_cast<double>(hparams.image_max_pixels) * max_pixels_tolerance;
+}
+
 mtmd_image_preprocessor_llava_uhd::slice_instructions mtmd_image_preprocessor_lfm2::get_slice_instructions(const clip_image_size & original_size) {
     mtmd_image_preprocessor_llava_uhd::slice_instructions inst;
     const int align_size = hparams.patch_size * hparams.n_merge;
     inst.overview_size = img_tool::calc_size_preserved_ratio(
                             original_size,
                             { align_size, hparams.image_min_pixels, hparams.image_max_pixels, 0 });
-    // tile if either dimension exceeds tile_size with tolerance
-    const bool needs_tiling = original_size.width > tile_size * max_pixels_tolerance || original_size.height > tile_size * max_pixels_tolerance;
+
+    const bool needs_tiling = should_tile(hparams, original_size);
 
     if (!needs_tiling) {
         inst.refined_size = clip_image_size{0, 0};

tools/mtmd/mtmd-image.h

@@ -148,6 +148,8 @@ struct mtmd_image_preprocessor_lfm2 : mtmd_image_preprocessor_llava_uhd {
     mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
     slice_instructions get_slice_instructions(const clip_image_size & original_size) override;
 
+    static bool should_tile(const clip_hparams & hparams, const clip_image_size & original_size);
+
 private:
     clip_image_size find_closest_aspect_ratio(
             float aspect_ratio,

tools/perplexity/perplexity.cpp

@@ -2023,7 +2023,6 @@ int llama_perplexity(int argc, char ** argv) {
     }
 
     const int32_t n_ctx = params.n_ctx;
-
     if (n_ctx <= 0) {
         LOG_ERR("%s: perplexity tool requires '--ctx-size' > 0\n", __func__);
         return 1;

tools/ui/src/lib/components/app/chat/ChatForm/ChatForm.svelte

@@ -48,6 +48,7 @@
 		containsFileMentionLink,
 		findCommandToken,
 		findMentionToken,
+		getConversationModel,
 		isIMEComposing,
 		isOffsetInCodeBlock,
 		parseClipboardContent,
@@ -190,31 +191,9 @@
 
 	let isRouter = $derived(serverStore.isRouterMode);
 	let conversationModel = $derived(
-		chatStore.getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
+		getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
 	);
-	let activeModelId = $derived.by(() => {
-		const options = modelsStore.models;
-
-		if (!isRouter) {
-			return options.length > 0 ? options[0].model : null;
-		}
-
-		const selectedId = modelsStore.selectedModelId;
-
-		if (selectedId) {
-			const model = options.find((m) => m.id === selectedId);
-
-			if (model) return model.model;
-		}
-
-		if (conversationModel) {
-			const model = options.find((m) => m.model === conversationModel);
-
-			if (model) return model.model;
-		}
-
-		return null;
-	});
+	let activeModelId = $derived(modelsStore.activeModelId);
 
 	let hasModelSelected = $derived(
 		!isRouter || !!conversationModel || !!modelsStore.selectedModelId

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionsAdd.svelte

@@ -2,10 +2,10 @@
 	import ChatFormActionAddButton from './ChatFormActionAddButton.svelte';
 	import ChatFormActionAddDropdown from './ChatFormActionAddDropdown.svelte';
 	import ChatFormActionAddSheet from './ChatFormActionAddSheet.svelte';
-	import { isMobile } from '$lib/stores';
+	import { deviceStore } from '$lib/stores';
 </script>
 
-{#if isMobile.current}
+{#if deviceStore.isMobile}
 	<ChatFormActionAddSheet>
 		{#snippet trigger({ disabled, onclick })}
 			<ChatFormActionAddButton {disabled} {onclick} />

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionModels.svelte

@@ -1,6 +1,7 @@
 <script lang="ts">
 	import { ModelsSelectorDropdown, ModelsSelectorSheet } from '$lib/components/app';
-	import { chatStore, conversationsStore, isMobile, modelsStore, serverStore } from '$lib/stores';
+	import { conversationsStore, deviceStore, modelsStore, serverStore } from '$lib/stores';
+	import { getConversationModel } from '$lib/utils';
 
 	interface Props {
 		disabled?: boolean;
@@ -30,7 +31,7 @@
 	let isOffline = $derived(!!serverStore.error);
 
 	let conversationModel = $derived(
-		chatStore.getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
+		getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
 	);
 
 	let lastSyncedConversationModel: string | null = null;
@@ -74,29 +75,7 @@
 		}
 	});
 
-	let activeModelId = $derived.by(() => {
-		const options = modelsStore.models;
-
-		if (!isRouter) {
-			return options.length > 0 ? options[0].model : null;
-		}
-
-		const selectedId = modelsStore.selectedModelId;
-
-		if (selectedId) {
-			const model = options.find((m) => m.id === selectedId);
-
-			if (model) return model.model;
-		}
-
-		if (conversationModel) {
-			const model = options.find((m) => m.model === conversationModel);
-
-			if (model) return model.model;
-		}
-
-		return null;
-	});
+	let activeModelId = $derived(modelsStore.activeModelId);
 
 	let modelPropsVersion = $state(0); // Used to trigger reactivity after fetch
 
@@ -170,7 +149,7 @@
 	}
 </script>
 
-{#if isMobile.current}
+{#if deviceStore.isMobile}
 	<ModelsSelectorSheet
 		disabled={disabled || isOffline}
 		bind:this={selectorModelRef}

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormContextGauge/ChatFormContextGauge.svelte

@@ -1,15 +1,14 @@
 <script lang="ts">
 	import ContextGaugeDial from './ContextGaugeDial.svelte';
-	import { useContextGauge } from '$lib/hooks/use-context-gauge.svelte';
 	import {
-		chatStore,
-		conversationsStore,
 		gaugeTriggerClick,
 		gaugeTriggerEnter,
 		gaugeTriggerKeydown,
 		gaugeTriggerLeave,
 		gaugeTriggerPointerDown
-	} from '$lib/stores';
+	} from './gauge-popup.svelte';
+	import { useContextGauge } from '$lib/hooks/use-context-gauge.svelte';
+	import { chatStore, conversationsStore } from '$lib/stores';
 	import { untrack } from 'svelte';
 
 	const gauge = useContextGauge();

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormContextGauge/ContextGaugeDetails.svelte

@@ -1,9 +1,9 @@
 <script lang="ts">
 	import ContextGaugeDetailRow from './ContextGaugeDetailRow.svelte';
+	import { gaugePopup } from './gauge-popup.svelte';
 	import { ChevronDown } from '@lucide/svelte';
 	import * as Collapsible from '$lib/components/ui/collapsible';
 	import { STATS_UNITS } from '$lib/constants';
-	import { gaugePopup } from '$lib/stores/context-gauge-popup.svelte';
 
 	interface Props {
 		currentRead: number;

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormContextGauge/ContextGaugePopup.svelte

@@ -2,8 +2,13 @@
 	import { colorLevelBgClass, colorLevelTextClass } from './context-gauge';
 	import ContextGaugeDetails from './ContextGaugeDetails.svelte';
 	import ContextGaugeLoadModel from './ContextGaugeLoadModel.svelte';
+	import {
+		gaugeCardEnter,
+		gaugeCardLeave,
+		gaugePopup,
+		gaugePopupClose
+	} from './gauge-popup.svelte';
 	import { useContextGauge } from '$lib/hooks/use-context-gauge.svelte';
-	import { gaugeCardEnter, gaugeCardLeave, gaugePopup, gaugePopupClose } from '$lib/stores';
 	import { formatParameters } from '$lib/utils/formatters';
 
 	const gauge = useContextGauge();

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormInput/ChatFormInputBasic.svelte

@@ -1,5 +1,5 @@
 <script lang="ts">
-	import { isMobile } from '$lib/stores';
+	import { deviceStore } from '$lib/stores';
 	import { autoResizeTextarea } from '$lib/utils';
 	import { onMount } from 'svelte';
 
@@ -37,7 +37,7 @@
 	}
 
 	export function focus() {
-		if (isMobile.current) return;
+		if (deviceStore.isMobile) return;
 
 		textareaElement?.focus({ preventScroll: true });
 	}

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormInput/ChatFormInputRich.svelte

@@ -1,7 +1,7 @@
 <script lang="ts">
 	import { CODE_BLOCK, CODE_TOKEN_ATTR, UI_DATA_ATTRS } from '$lib/constants';
 	import { BooleanString, ChatFormInputRichTokenKind, ColorMode } from '$lib/enums';
-	import { isMobile } from '$lib/stores';
+	import { deviceStore } from '$lib/stores';
 	import type { ChatFormInputRichToken } from '$lib/types';
 	import type { SourceHistoryEntry } from '$lib/utils';
 	import {
@@ -750,7 +750,7 @@
 		syncEmptyState();
 		document.addEventListener('selectionchange', handleSelectionChange);
 
-		if (!isMobile.current) {
+		if (!deviceStore.isMobile) {
 			rootElement?.focus({ preventScroll: true });
 		}
 	});
@@ -792,7 +792,7 @@
 	}
 
 	export function focus() {
-		if (isMobile.current) return;
+		if (deviceStore.isMobile) return;
 
 		rootElement?.focus({ preventScroll: true });
 	}

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormPickers/ChatFormPickerMention.svelte

@@ -8,7 +8,7 @@
 	import { BuiltInTool, FileMentionEntryType, GlobSearchType, KeyboardKey } from '$lib/enums';
 	import { useDebouncedSearch } from '$lib/hooks/use-debounced-search.svelte';
 	import { usePickerNavigation } from '$lib/hooks/use-picker-navigation.svelte';
-	import { isMobile, settingsStore, toolsStore } from '$lib/stores';
+	import { deviceStore, settingsStore, toolsStore } from '$lib/stores';
 	import type { FileMentionEntry, GlobEntryResult } from '$lib/types';
 	import { abbreviateHome, runGlobSearchWithChildren } from '$lib/utils';
 
@@ -130,7 +130,7 @@
 		return searchError ? `Search failed - ${searchError}` : 'No matching files or folders';
 	});
 
-	const showTooltip = $derived(!isMobile.current);
+	const showTooltip = $derived(!deviceStore.isMobile);
 
 	$effect(() => {
 		if (typeof window === 'undefined') return;

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessage.svelte

@@ -11,7 +11,7 @@
 	import { setChatMessageActionsContext, setChatMessageEditContext } from '$lib/contexts';
 	import { AgenticSectionType, AttachmentType, MessageRole } from '$lib/enums';
 	import { DatabaseService } from '$lib/services/database.service';
-	import { chatStore, conversationsStore, isMobile } from '$lib/stores';
+	import { chatStore, conversationsStore, deviceStore } from '$lib/stores';
 	import type {
 		ChatMessageActions,
 		ChatMessageDeletionInfo,
@@ -304,7 +304,7 @@
 
 	// After the system message flow ends, hand focus to the main chat form
 	function focusMainChatForm() {
-		if (isMobile.current) return;
+		if (deviceStore.isMobile) return;
 
 		document.querySelector<HTMLTextAreaElement>('.chat-screen-form-wrapper textarea')?.focus();
 	}

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte

@@ -21,8 +21,7 @@
 	import {
 		chatStore,
 		conversationsStore,
-		device,
-		isMobile,
+		deviceStore,
 		serverStore,
 		settingsStore
 	} from '$lib/stores';
@@ -32,7 +31,7 @@
 	let { showCenteredEmpty = false } = $props();
 
 	let disableAutoScroll = $derived(
-		Boolean(settingsStore.config.disableAutoScroll) || isMobile.current
+		Boolean(settingsStore.config.disableAutoScroll) || deviceStore.isMobile
 	);
 	let isMobileUserScrolledUp = $state(false);
 	let mobileScrollDownHint = $state(false);
@@ -52,11 +51,11 @@
 	let hasPropsError = $derived(!!serverStore.error);
 	let isCurrentConversationLoading = $derived(chatStore.isLoading || chatStore.isStreaming());
 	let chatFormBottomPosition = $derived.by(() => {
-		if (!isMobile.current) return '1rem';
+		if (!deviceStore.isMobile) return '1rem';
 
-		if (device.isStandalone) return '1.5rem';
+		if (deviceStore.isStandalone) return '1.5rem';
 
-		if (device.isIOSSafari) return '0.25rem';
+		if (deviceStore.isIOSSafari) return '0.25rem';
 
 		return '0.5rem';
 	});
@@ -84,7 +83,7 @@
 	});
 
 	function handleMobileScroll() {
-		if (!isMobile.current) return;
+		if (!deviceStore.isMobile) return;
 
 		const container = scroll.chatScrollContainer;
 
@@ -184,7 +183,7 @@
 	}
 
 	function handleSendLikeScroll() {
-		if (!isMobile.current) {
+		if (!deviceStore.isMobile) {
 			autoScroll.enable();
 		}
 
@@ -197,7 +196,7 @@
 				'.chat-message:nth-last-child(2) .chat-message-user .chat-message-user-bubble'
 			) as HTMLElement | null;
 
-			if (isMobile.current) {
+			if (deviceStore.isMobile) {
 				// Keep the last user message bubble just above the input on mobile
 				const bubbleHeight = lastUserBubble?.scrollHeight ?? 0;
 				const baseHeight = container.scrollHeight - innerHeight;
@@ -220,7 +219,7 @@
 			}
 		}, 100);
 
-		if (isMobile.current) {
+		if (deviceStore.isMobile) {
 			autoScroll.setDisabled(disableAutoScroll);
 			mobileScrollDownHint = true;
 			mobileScrollDownHintLockedUntil = Date.now() + 500;
@@ -243,7 +242,8 @@
 
 	$effect(() => {
 		const shouldDisableAutoScroll =
-			settingsStore.config.disableAutoScroll || (isMobile.current && isCurrentConversationLoading);
+			settingsStore.config.disableAutoScroll ||
+			(deviceStore.isMobile && isCurrentConversationLoading);
 
 		autoScroll.setDisabled(shouldDisableAutoScroll);
 
@@ -266,7 +266,7 @@
 			autoScroll.enable();
 		}
 
-		if (isMobile.current && isCurrentConversationLoading) {
+		if (deviceStore.isMobile && isCurrentConversationLoading) {
 			mobileScrollDownHint = true;
 			mobileScrollDownHintLockedUntil = Date.now() + 500;
 		}
@@ -318,9 +318,9 @@
 		<div
 			class={[
 				'pointer-events-none md:sticky fixed  mt-auto transition-all duration-200',
-				device.isStandalone
+				deviceStore.isStandalone
 					? 'bottom-6 right-4 left-4'
-					: device.isIOSSafari
+					: deviceStore.isIOSSafari
 						? 'bottom-1 left-2 right-2'
 						: 'bottom-2 right-2 left-2',
 				isEmpty ? 'md:bottom-[calc(50dvh-7rem)] 2xl:bottom-[calc(50dvh-4rem)]' : 'md:bottom-4'
@@ -336,7 +336,7 @@
 			{/if}
 
 			<div class="pointer-events-none flex flex-col gap-6 items-center w-full">
-				{#if (isMobile.current ? mobileScrollDownHint || isMobileUserScrolledUp : autoScroll.userScrolledUp) && page.url.hash.includes(ROUTES.CHAT) && page.params.id}
+				{#if (deviceStore.isMobile ? mobileScrollDownHint || isMobileUserScrolledUp : autoScroll.userScrolledUp) && page.url.hash.includes(ROUTES.CHAT) && page.params.id}
 					<ChatScreenActionScrollDown
 						onclick={() => {
 							mobileScrollDownHint = false;

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenForm.svelte

@@ -3,7 +3,7 @@
 	import { page } from '$app/state';
 	import { ChatForm } from '$lib/components/app';
 	import { useDraftMessages } from '$lib/hooks/use-draft-messages.svelte';
-	import { isMobile } from '$lib/stores';
+	import { deviceStore } from '$lib/stores';
 	import { onMount } from 'svelte';
 
 	interface Props {
@@ -120,13 +120,13 @@
 	}
 
 	onMount(() => {
-		if (!isMobile.current) {
+		if (!deviceStore.isMobile) {
 			setTimeout(focusFormUnlessCaptured, 100);
 		}
 	});
 
 	afterNavigate((navigation) => {
-		if (navigation?.from != null && !isMobile.current) {
+		if (navigation?.from != null && !deviceStore.isMobile) {
 			setTimeout(focusFormUnlessCaptured, 100);
 		}
 	});

tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigation.svelte

@@ -14,7 +14,7 @@
 	import { useKeyboardShortcuts } from '$lib/hooks/use-keyboard-shortcuts.svelte';
 	import { useMarqueeSelection } from '$lib/hooks/use-marquee-selection.svelte';
 	import { RouterService } from '$lib/services/router.service';
-	import { chatStore, conversationsStore, device, isMobile, settingsStore } from '$lib/stores';
+	import { chatStore, conversationsStore, deviceStore, settingsStore } from '$lib/stores';
 	import { buildConversationTree } from '$lib/utils';
 	import { circIn } from 'svelte/easing';
 	import { SvelteSet } from 'svelte/reactivity';
@@ -36,7 +36,7 @@
 	let logoHovered = $state(false);
 
 	const isStripExpanded = $derived(isExpandedMode || hoveredTooltip !== null);
-	const isOnMobile = $derived(isMobile.current);
+	const isOnMobile = $derived(deviceStore.isMobile);
 	const alwaysShowOnDesktop = $derived(settingsStore.config.alwaysShowSidebarOnDesktop as boolean);
 
 	$effect(() => {
@@ -65,7 +65,7 @@
 	});
 
 	$effect(() => {
-		if (isMobile.current && page.url.hash.includes(ROUTES.SEARCH)) {
+		if (deviceStore.isMobile && page.url.hash.includes(ROUTES.SEARCH)) {
 			isExpandedMode = false;
 		}
 	});
@@ -227,7 +227,7 @@
 	}
 
 	async function selectConversation(id: string) {
-		if (isMobile.current) {
+		if (deviceStore.isMobile) {
 			scheduleMobileCollapse();
 		}
 
@@ -315,9 +315,9 @@
 			'fixed md:sticky top-2 left-2 md:left-0 md:ml-2 md:mt-2 pt-2 z-10 w-[calc(100dvw-1rem)]',
 			'md:h-[calc(100dvh-1.125rem)]',
 			isExpandedMode &&
-				(device.isStandalone
+				(deviceStore.isStandalone
 					? 'h-[calc(100dvh-2rem)]'
-					: device.isIOSDevice
+					: deviceStore.isIOSDevice
 						? 'h-[calc(100dvh-0.5rem)]'
 						: 'h-[calc(100dvh-1rem)]'),
 			'rounded-3xl md:rounded-2xl',
@@ -353,7 +353,7 @@
 
 			{#if isOnMobile || (isExpandedMode && !alwaysShowOnDesktop)}
 				<div
-					class="flex items-center transition-all duration-150 ease-out {isMobile.current &&
+					class="flex items-center transition-all duration-150 ease-out {deviceStore.isMobile &&
 					!isExpandedMode
 						? 'opacity-0 h-0!'
 						: ''}"
@@ -361,7 +361,7 @@
 					out:fade={{ duration: 100 }}
 				>
 					<ActionIcon
-						icon={isMobile.current ? X : PanelLeftClose}
+						icon={deviceStore.isMobile ? X : PanelLeftClose}
 						size="lg"
 						iconSize="h-4.5 w-4.5 md:h-4 md:w-4"
 						class="backdrop-blur-none md:h-9 md:w-9 h-10 w-10 rounded-full mr-1 hover:bg-accent!"
@@ -375,9 +375,9 @@
 		</div>
 
 		<div
-			class="mt-2 flex min-h-0 flex-1 flex-col gap-4 md:gap-1 {isMobile.current
+			class="mt-2 flex min-h-0 flex-1 flex-col gap-4 md:gap-1 {deviceStore.isMobile
 				? 'transition-[opacity,height] duration-200 ease-out'
-				: ''} {isMobile.current && !isExpandedMode ? 'opacity-0 !h-0' : ''}"
+				: ''} {deviceStore.isMobile && !isExpandedMode ? 'opacity-0 !h-0' : ''}"
 			in:fade={{ duration: 200 }}
 			out:fade={{ duration: 200 }}
 		>
@@ -395,7 +395,7 @@
 					isSearchModeActive = true;
 				}}
 				onNewChat={() => {
-					if (isMobile.current) {
+					if (deviceStore.isMobile) {
 						scheduleMobileCollapse();
 					}
 				}}

tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigationActions.svelte

@@ -12,7 +12,7 @@
 		SIDEBAR_ACTIONS_ITEMS
 	} from '$lib/constants';
 	import { TooltipSide } from '$lib/enums';
-	import { isMobile } from '$lib/stores';
+	import { deviceStore } from '$lib/stores';
 	import type { Component } from 'svelte';
 	import { onMount } from 'svelte';
 	import { circIn } from 'svelte/easing';
@@ -42,7 +42,7 @@
 	let showIcons = $state(false);
 	let searchInputRef = $state<HTMLInputElement | null>(null);
 
-	const isOnMobile = $derived(isMobile.current);
+	const isOnMobile = $derived(deviceStore.isMobile);
 
 	$effect(() => {
 		if (isSearchModeActive && searchInputRef) {
@@ -107,7 +107,7 @@
 	>
 		{#each SIDEBAR_ACTIONS_ITEMS as item, i (item.tooltip)}
 			{@const isActive = isItemActive(item)}
-			{@const isSearchOnMobile = item.icon === Search && isMobile.current}
+			{@const isSearchOnMobile = item.icon === Search && deviceStore.isMobile}
 			{@const itemHref = isSearchOnMobile ? ROUTES.SEARCH : item.route}
 			{@const itemOnClick = item.route
 				? () => {
@@ -156,7 +156,7 @@
 	<div class="{className} flex-col gap-1 hidden md:flex">
 		{#each SIDEBAR_ACTIONS_ITEMS as item, i (item.tooltip)}
 			{@const isActive = isItemActive(item)}
-			{@const isSearchOnMobile = item.icon === Search && isMobile.current}
+			{@const isSearchOnMobile = item.icon === Search && deviceStore.isMobile}
 			{@const itemOnClick = item.route
 				? () => {
 						onNewChat?.();

tools/ui/src/lib/components/app/settings/SettingsChat/SettingsChatImportExportTab.svelte

@@ -8,6 +8,7 @@
 	} from '$lib/components/app';
 	import SettingsGroup from '$lib/components/app/settings/SettingsGroup.svelte';
 	import { ConversationSelectionMode, FileExtensionText, HtmlInputType } from '$lib/enums';
+	import { ConversationTransferService } from '$lib/services';
 	import { conversationsStore, settingsStore } from '$lib/stores';
 	import { createMessageCountMap } from '$lib/utils';
 	import { fade } from 'svelte/transition';
@@ -147,9 +148,9 @@
 			);
 
 			if (allData.length === 1) {
-				conversationsStore.downloadConversationFile(allData[0]);
+				ConversationTransferService.downloadConversationFile(allData[0]);
 			} else {
-				conversationsStore.downloadConversationsArchive(allData);
+				ConversationTransferService.downloadConversationsArchive(allData);
 			}
 
 			exportedConversations = selectedConversations;
@@ -177,7 +178,7 @@
 				if (!file) return;
 
 				try {
-					const importedData = await conversationsStore.parseImportFile(file);
+					const importedData = await ConversationTransferService.parseImportFile(file);
 
 					if (importedData.length === 0) {
 						throw new Error('No conversations found in file');

tools/ui/src/lib/constants/api-endpoints.constants.ts

@@ -31,5 +31,11 @@ export const API_STREAM = {
 	LOOKUP: './v1/streams/lookup'
 };
 
+// query params for the resumable stream routes
+export const STREAM_QUERY_PARAMS = {
+	CONV_ID: 'conv_id',
+	FROM: 'from'
+} as const;
+
 /** CORS proxy endpoint path */
 export const CORS_PROXY_ENDPOINT = '/cors-proxy';

tools/ui/src/lib/hooks/use-chat-screen-active-model.svelte.ts

@@ -8,36 +8,15 @@
  * demand if they aren't cached yet.
  */
 
-import { chatStore, conversationsStore, modelsStore, serverStore } from '$lib/stores';
+import { conversationsStore, modelsStore, serverStore } from '$lib/stores';
+import { getConversationModel } from '$lib/utils';
 
 export function useChatScreenActiveModel() {
 	const isRouter = $derived(serverStore.isRouterMode);
 	const conversationModel = $derived(
-		chatStore.getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
+		getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
 	);
-	const activeModelId = $derived.by(() => {
-		const options = modelsStore.models;
-
-		if (!isRouter) {
-			return options.length > 0 ? options[0].model : null;
-		}
-
-		const selectedId = modelsStore.selectedModelId;
-
-		if (selectedId) {
-			const model = options.find((m) => m.id === selectedId);
-
-			if (model) return model.model;
-		}
-
-		if (conversationModel) {
-			const model = options.find((m) => m.model === conversationModel);
-
-			if (model) return model.model;
-		}
-
-		return null;
-	});
+	const activeModelId = $derived(modelsStore.activeModelId);
 
 	let modelPropsVersion = $state(0);
 

tools/ui/src/lib/hooks/use-pwa.svelte.ts

@@ -58,7 +58,7 @@ export function usePwa() {
 		// PWA pages update via the service worker path; the storage check is the non-PWA fallback only
 		if (navigator.serviceWorker?.controller) return;
 
-		const currentVersion = versionStore.value;
+		const currentVersion = versionStore.frontend;
 
 		if (!currentVersion) return;
 

tools/ui/src/lib/hooks/use-reasoning-menu.svelte.ts

@@ -1,8 +1,9 @@
 import { REASONING_EFFORT_LEVELS, REASONING_EFFORT_TOKENS } from '$lib/constants';
 import { ReasoningEffort } from '$lib/enums';
-import { chatStore, conversationsStore, modelsStore, serverStore } from '$lib/stores';
+import { conversationsStore, modelsStore, serverStore } from '$lib/stores';
 import type { ReasoningEffortLevel } from '$lib/types';
 import type { DatabaseMessage } from '$lib/types/database';
+import { getConversationModel } from '$lib/utils';
 
 export interface UseReasoningMenuReturn {
 	readonly modelSupportsThinking: boolean;
@@ -24,7 +25,7 @@ export interface UseReasoningMenuReturn {
  */
 export function useReasoningMenu(): UseReasoningMenuReturn {
 	const conversationModel = $derived(
-		chatStore.getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
+		getConversationModel(conversationsStore.activeMessages as DatabaseMessage[])
 	);
 	// a router chat can carry reasoning from an earlier turn before the props
 	// cache is primed, so a model that already produced thinking still qualifies

tools/ui/src/lib/services/chat.service.ts

@@ -13,6 +13,7 @@ import {
 	SSE_DATA_PREFIX,
 	SSE_DONE_MARKER,
 	SSE_LINE_SEPARATOR,
+	STREAM_QUERY_PARAMS,
 	STREAM_RESUME_LOCALSTORAGE_KEY_PREFIX,
 	STREAM_VISIBILITY_KICK_MS
 } from '$lib/constants';
@@ -33,6 +34,7 @@ import type {
 	ApiStreamSession
 } from '$lib/types/api';
 import { isAbortError } from '$lib/utils/abort';
+import { ApiError } from '$lib/utils/api-fetch';
 import { getAuthHeaders, getJsonHeaders } from '$lib/utils/api-headers';
 import { formatAttachmentText } from '$lib/utils/formatters';
 import { streamIdentity } from '$lib/utils/stream-identity';
@@ -529,7 +531,7 @@ export class ChatService {
 		try {
 			const id = streamIdentity(conversationId, model);
 
-			await fetch(`${API_STREAM.BASE}?conv_id=${encodeURIComponent(id)}`, {
+			await fetch(ChatService.buildStreamUrl(id), {
 				headers: getAuthHeaders(),
 				method: 'DELETE'
 			});
@@ -538,6 +540,46 @@ export class ChatService {
 		}
 	}
 
+	/**
+	 * Look up server-side stream sessions for the given conversation ids. Ids carry the frozen
+	 * conv::model identity when a model was bound at POST time.
+	 */
+	static async lookupStreamSessions(conversationIds: string[]): Promise<ApiStreamSession[]> {
+		const resp = await fetch(API_STREAM.LOOKUP, {
+			body: JSON.stringify({ conversation_ids: conversationIds }),
+			headers: getJsonHeaders(),
+			method: 'POST'
+		});
+
+		if (!resp.ok) {
+			throw new ApiError(`Stream lookup failed with HTTP ${resp.status}`, resp.status);
+		}
+
+		const body = (await resp.json()) as unknown;
+
+		if (!Array.isArray(body)) {
+			throw new Error('Stream lookup returned a non-array response');
+		}
+
+		return body as ApiStreamSession[];
+	}
+
+	/**
+	 * Fetch the full replay of a server-side stream from byte 0. Returns the raw Response so the
+	 * caller can pipe it through the SSE parser like a fresh stream.
+	 */
+	static async fetchStreamReplay(streamId: string): Promise<Response> {
+		const resp = await fetch(ChatService.buildStreamUrl(streamId, 0), {
+			headers: getAuthHeaders()
+		});
+
+		if (!resp.ok) {
+			throw new ApiError(`Stream replay failed with HTTP ${resp.status}`, resp.status);
+		}
+
+		return resp;
+	}
+
 	/**
 	 * Pick the running session to splice into when discoverActiveStream lists candidates for a
 	 * conversation. Finalized sessions are not candidates: their final content was already written
@@ -629,6 +671,15 @@ export class ChatService {
 		return streamIdentity(conversationId, model);
 	}
 
+	// build the replay route url for a stream identity, from is the resume byte offset, omitted
+	// for the cancel route
+	private static buildStreamUrl(streamId: string, from?: number): string {
+		const query = `${STREAM_QUERY_PARAMS.CONV_ID}=${encodeURIComponent(streamId)}`;
+		const offset = from === undefined ? '' : `&${STREAM_QUERY_PARAMS.FROM}=${from}`;
+
+		return `${API_STREAM.BASE}?${query}${offset}`;
+	}
+
 	/**
 	 * Reconnect to an interrupted stream for this conversation. Returns the fetch Response so the
 	 * existing SSE parser drains it like a fresh stream. The server returns 200 on success, 404 if
@@ -642,13 +693,10 @@ export class ChatService {
 		const ac = new AbortController();
 
 		try {
-			const resp = await fetch(
-				`${API_STREAM.BASE}?conv_id=${encodeURIComponent(streamId)}&from=0`,
-				{
-					headers: getAuthHeaders(),
-					signal: ac.signal
-				}
-			);
+			const resp = await fetch(ChatService.buildStreamUrl(streamId, 0), {
+				headers: getAuthHeaders(),
+				signal: ac.signal
+			});
 
 			ac.abort();
 
@@ -668,7 +716,7 @@ export class ChatService {
 		const state = ChatService.getStreamState(conversationId);
 		const from = state?.bytesReceived ?? 0;
 		const id = streamIdentity(conversationId, model);
-		const url = `${API_STREAM.BASE}?conv_id=${encodeURIComponent(id)}&from=${from}`;
+		const url = ChatService.buildStreamUrl(id, from);
 
 		return await fetch(url, { headers: getAuthHeaders(), method: 'GET', signal });
 	}

tools/ui/src/lib/services/conversation-transfer.service.ts

@@ -0,0 +1,279 @@
+/**
+ * ConversationTransferService - Stateless conversation import/export layer
+ *
+ * Owns the session file format (one JSONL record per line: a SESSION header
+ * followed by MESSAGE records), ZIP archiving and browser downloads.
+ * DB access and store refreshes stay in conversationsStore.
+ */
+
+import { EXPORT_CONV, NEWLINE, ZIP_MAGIC } from '$lib/constants';
+import {
+	FileExtensionText,
+	MimeTypeApplication,
+	MimeTypeText,
+	SessionRecordType
+} from '$lib/enums';
+import { strFromU8, strToU8, unzipSync, zipSync } from 'fflate';
+
+export class ConversationTransferService {
+	/**
+	 *
+	 *
+	 * JSONL Session Format
+	 *
+	 *
+	 */
+
+	/**
+	 * Serializes a session (a conversation with its messages) as JSONL.
+	 * The first line is the session header (a `SessionRecordType.SESSION` record
+	 * carrying the conversation properties); each subsequent line is a single message.
+	 * @param data - The exported conversation payload
+	 * @returns The JSONL string (one record per line)
+	 */
+	static serializeSessionToJsonl(data: ExportedConversation): string {
+		const { conv, messages } = data;
+		const sessionLine = JSON.stringify({
+			harness: EXPORT_CONV.HARNESS,
+			type: SessionRecordType.SESSION,
+			...conv
+		});
+		const messageLines = messages.map((message: DatabaseMessage) => {
+			// `toolCalls` is stored as a JSON string; drop it when empty, otherwise parse it.
+			const { toolCalls, ...rest } = message;
+			const normalized = toolCalls ? { ...rest, toolCalls: JSON.parse(toolCalls) } : rest;
+
+			return JSON.stringify({ message: normalized, type: SessionRecordType.MESSAGE });
+		});
+
+		return [sessionLine, ...messageLines].join(NEWLINE);
+	}
+
+	/**
+	 * Parses the JSONL session format produced by {@link serializeSessionToJsonl}.
+	 * A `SessionRecordType.SESSION` line starts a new session; following
+	 * `SessionRecordType.MESSAGE` lines are appended to it. Supports multiple
+	 * sessions in a single file.
+	 * @param text - The JSONL file contents
+	 * @returns The parsed conversations with their messages
+	 */
+	static parseSessionsJsonl(text: string): ExportedConversation[] {
+		const sessions: ExportedConversation[] = [];
+
+		let current: ExportedConversation | null = null;
+
+		for (const line of text.split(NEWLINE)) {
+			const trimmed = line.trim();
+
+			if (!trimmed) continue;
+
+			const record = JSON.parse(trimmed);
+
+			if (record.type === SessionRecordType.SESSION) {
+				// Drop the discriminator and harness marker; the rest is the conversation.
+				const conv = { ...record };
+
+				delete conv.type;
+				delete conv.harness;
+				current = { conv: conv as DatabaseConversation, messages: [] };
+				sessions.push(current);
+			} else if (record.type === SessionRecordType.MESSAGE) {
+				if (!current) {
+					throw new Error('Invalid JSONL: message record before any session record');
+				}
+
+				const message = record.message as DatabaseMessage;
+
+				// `toolCalls` is parsed to an array on export; the DB stores it as a string.
+				if (message.toolCalls !== undefined && typeof message.toolCalls !== 'string') {
+					message.toolCalls = JSON.stringify(message.toolCalls);
+				}
+
+				current.messages.push(message);
+			}
+			// Ignore unknown record types for forward compatibility.
+		}
+
+		return sessions;
+	}
+
+	/**
+	 * Reports whether the text is the JSONL session format, whose first non-empty
+	 * line is a `SessionRecordType.SESSION` record. A legacy JSON export starts
+	 * with an array or an object that has no such discriminator.
+	 * @param text - The file contents
+	 */
+	private static isSessionsJsonl(text: string): boolean {
+		const trimmed = text.trimStart();
+		const lineEnd = trimmed.indexOf(NEWLINE);
+		const firstLine = lineEnd === -1 ? trimmed : trimmed.slice(0, lineEnd);
+
+		try {
+			return JSON.parse(firstLine).type === SessionRecordType.SESSION;
+		} catch {
+			// Not a standalone JSON record, so not the JSONL format.
+			return false;
+		}
+	}
+
+	/**
+	 * Parses an import file into conversations, accepting the current JSONL and
+	 * ZIP formats as well as the legacy JSON format. The format comes from the
+	 * contents, so an import works whatever the file is named.
+	 * @param file - The user-selected file
+	 * @returns The parsed conversations with their messages
+	 */
+	static async parseImportFile(file: File): Promise<ExportedConversation[]> {
+		const bytes = new Uint8Array(await file.arrayBuffer());
+
+		if (ZIP_MAGIC.every((byte, index) => bytes[index] === byte)) {
+			const entries = unzipSync(bytes);
+			const sessions: ExportedConversation[] = [];
+
+			for (const [entryName, entryBytes] of Object.entries(entries)) {
+				if (!entryName.toLowerCase().endsWith(FileExtensionText.JSONL)) continue;
+
+				sessions.push(...ConversationTransferService.parseSessionsJsonl(strFromU8(entryBytes)));
+			}
+
+			return sessions;
+		}
+
+		const text = strFromU8(bytes);
+
+		if (ConversationTransferService.isSessionsJsonl(text)) {
+			return ConversationTransferService.parseSessionsJsonl(text);
+		}
+
+		// Legacy JSON format: an array of conversations or a single conversation object.
+		const parsed = JSON.parse(text);
+
+		if (Array.isArray(parsed)) {
+			return parsed;
+		}
+
+		if (parsed && typeof parsed === 'object' && 'conv' in parsed && 'messages' in parsed) {
+			return [parsed];
+		}
+
+		throw new Error(
+			'Invalid file format: expected array of conversations or single conversation object'
+		);
+	}
+
+	/**
+	 *
+	 *
+	 * Downloads
+	 *
+	 *
+	 */
+
+	/**
+	 * Generates a sanitized filename for a conversation export
+	 * @param conversation - The conversation metadata
+	 * @param msgs - Optional array of messages belonging to the conversation
+	 * @returns The generated filename string
+	 */
+	static generateConversationFilename(
+		conversation: { id?: string; name?: string },
+		msgs?: DatabaseMessage[]
+	): string {
+		const conversationName = (conversation.name ?? '').trim().toLowerCase();
+		const sanitizedName = conversationName
+			.replace(EXPORT_CONV.NON_ALPHANUMERIC_REGEX, EXPORT_CONV.NONALNUM_REPLACEMENT)
+			.replace(EXPORT_CONV.MULTIPLE_UNDERSCORE_REGEX, '_')
+			.substring(0, EXPORT_CONV.NAME_SUFFIX_MAX_LENGTH);
+		// If we have messages, use the timestamp of the newest message
+		const referenceDate = msgs?.length
+			? new Date(Math.max(...msgs.map((m) => m.timestamp)))
+			: new Date();
+		const iso = referenceDate.toISOString().slice(0, EXPORT_CONV.ISO_TIMESTAMP_SLICE);
+		const formattedDate = iso
+			.replace(EXPORT_CONV.ISO_DATE_TIME_SEPARATOR, EXPORT_CONV.ISO_DATE_TIME_SEPARATOR_REPLACEMENT)
+			.replaceAll(EXPORT_CONV.ISO_TIME_SEPARATOR, EXPORT_CONV.ISO_TIME_SEPARATOR_REPLACEMENT);
+		const trimmedConvId = conversation.id?.slice(0, EXPORT_CONV.ID_TRIM_LENGTH) ?? '';
+
+		return `${formattedDate}_conv_${trimmedConvId}_${sanitizedName}${FileExtensionText.JSONL}`;
+	}
+
+	/**
+	 * Triggers a browser download of the provided exported conversation data
+	 * @param data - The exported conversation payload (a single conversation with its messages)
+	 * @param filename - Filename; if omitted, a deterministic name is generated
+	 */
+	static downloadConversationFile(data: ExportedConversation, filename?: string): void {
+		const { conv: conversation, messages: msgs } = data;
+
+		if (!conversation) {
+			console.error('Invalid data: missing conversation');
+
+			return;
+		}
+
+		const downloadFilename =
+			filename ?? ConversationTransferService.generateConversationFilename(conversation, msgs);
+		const jsonl = ConversationTransferService.serializeSessionToJsonl(data);
+		const blob = new Blob([jsonl], { type: MimeTypeText.JSONL });
+
+		ConversationTransferService.triggerDownload(blob, downloadFilename);
+	}
+
+	/**
+	 * Triggers a browser download of multiple conversations as a `.zip`, one
+	 * `.jsonl` file per conversation.
+	 * @param data - The conversations to export
+	 */
+	static downloadConversationsArchive(data: ExportedConversation[]): void {
+		if (data.length === 0) {
+			console.error('Invalid data: no conversations to export');
+
+			return;
+		}
+
+		const usedNames = new Set<string>();
+		const files: Record<string, Uint8Array> = {};
+
+		for (const session of data) {
+			const baseName = ConversationTransferService.generateConversationFilename(
+				session.conv,
+				session.messages
+			);
+
+			// Disambiguate any duplicate filenames within the archive.
+			let entryName = baseName;
+			let suffix = 1;
+
+			while (usedNames.has(entryName)) {
+				entryName = baseName.replace(
+					new RegExp(`${FileExtensionText.JSONL}$`),
+					`_${suffix++}${FileExtensionText.JSONL}`
+				);
+			}
+			usedNames.add(entryName);
+
+			files[entryName] = strToU8(ConversationTransferService.serializeSessionToJsonl(session));
+		}
+
+		const archiveName = `${new Date().toISOString().split(EXPORT_CONV.ISO_DATE_TIME_SEPARATOR)[0]}_conversations${FileExtensionText.ZIP}`;
+		const zipped = zipSync(files);
+		const blob = new Blob([zipped], { type: MimeTypeApplication.ZIP });
+
+		ConversationTransferService.triggerDownload(blob, archiveName);
+	}
+
+	/**
+	 * Triggers a browser download of a blob under the given filename.
+	 */
+	private static triggerDownload(blob: Blob, filename: string): void {
+		const url = URL.createObjectURL(blob);
+		const a = document.createElement('a');
+
+		a.href = url;
+		a.download = filename;
+		document.body.appendChild(a);
+		a.click();
+		document.body.removeChild(a);
+		URL.revokeObjectURL(url);
+	}
+}

tools/ui/src/lib/services/index.ts

@@ -103,6 +103,15 @@ export { ChatService } from './chat.service';
  */
 export { DatabaseService } from './database.service';
 
+/**
+ * **ConversationTransferService** - Conversation import/export format layer
+ *
+ * Owns the JSONL session format (SESSION header + MESSAGE records), ZIP
+ * archiving and browser downloads. Stateless; DB access and store refreshes
+ * stay in conversationsStore.
+ */
+export { ConversationTransferService } from './conversation-transfer.service';
+
 /**
  * **ModelsService** - Model management API communication
  *

tools/ui/src/lib/services/models.service.ts

@@ -1,7 +1,15 @@
-import { API_MODELS, MODEL_ID } from '$lib/constants';
+import { base } from '$app/paths';
+import {
+	API_MODELS,
+	MODEL_ID,
+	SSE_DATA_PREFIX,
+	SSE_LINE_SEPARATOR,
+	SSE_RECORD_SEPARATOR
+} from '$lib/constants';
 import { ServerModelStatus } from '$lib/enums';
 import type { ParsedModelId } from '$lib/types/models';
 import { apiFetch, apiPost, normalizeModelName } from '$lib/utils';
+import { getAuthHeaders } from '$lib/utils/api-headers';
 
 export class ModelsService {
 	/**
@@ -100,6 +108,89 @@ export class ModelsService {
 		return model.status.value === ServerModelStatus.LOADING;
 	}
 
+	/**
+	 *
+	 *
+	 * Status Feed
+	 *
+	 *
+	 */
+
+	private static readonly SSE_RECONNECT_MS = 1000;
+
+	/**
+	 * Read the /models/sse feed and invoke onEvent for each parsed envelope.
+	 * Reconnects on network drops until the signal aborts. Splits the byte
+	 * stream into SSE records on the blank line boundary; the payload rides in
+	 * the data lines as a JSON envelope with its own model, event and data fields.
+	 */
+	static async watchModelEvents(
+		signal: AbortSignal,
+		onEvent: (event: ApiModelsSseEvent) => void
+	): Promise<void> {
+		const decoder = new TextDecoder();
+
+		while (!signal.aborted) {
+			try {
+				const response = await fetch(`${base}${API_MODELS.SSE}`, {
+					headers: getAuthHeaders(),
+					signal
+				});
+
+				if (response.ok && response.body) {
+					const reader = response.body.getReader();
+
+					let buffer = '';
+
+					while (!signal.aborted) {
+						const { done, value } = await reader.read();
+
+						if (done) break;
+
+						buffer += decoder.decode(value, { stream: true });
+
+						let boundary = buffer.indexOf(SSE_RECORD_SEPARATOR);
+
+						while (boundary !== -1) {
+							const event = ModelsService.parseStatusRecord(buffer.slice(0, boundary));
+
+							if (event) onEvent(event);
+
+							buffer = buffer.slice(boundary + SSE_RECORD_SEPARATOR.length);
+							boundary = buffer.indexOf(SSE_RECORD_SEPARATOR);
+						}
+					}
+				}
+			} catch {
+				// network drop or abort falls through to the reconnect delay
+			}
+
+			if (signal.aborted) return;
+
+			await new Promise((resolve) => setTimeout(resolve, ModelsService.SSE_RECONNECT_MS));
+		}
+	}
+
+	/**
+	 * Parse one SSE record into its JSON envelope, or null when the record
+	 * carries no data payload or malformed JSON.
+	 */
+	private static parseStatusRecord(record: string): ApiModelsSseEvent | null {
+		const payload = record
+			.split(SSE_LINE_SEPARATOR)
+			.filter((line) => line.startsWith(SSE_DATA_PREFIX))
+			.map((line) => line.slice(SSE_DATA_PREFIX.length).trim())
+			.join(SSE_LINE_SEPARATOR);
+
+		if (payload.length === 0) return null;
+
+		try {
+			return JSON.parse(payload) as ApiModelsSseEvent;
+		} catch {
+			return null;
+		}
+	}
+
 	/**
 	 *
 	 *

tools/ui/src/lib/stores/build-info.svelte.ts

@@ -1,45 +0,0 @@
-/**
- * buildInfoStore - llama.cpp build information
- *
- * Reads the build version from `build.json` — embedded at llama.cpp build time
- * with the llama.cpp build number (LLAMA_BUILD_NUMBER). Shown in the UI when
- * `showBuildVersion` is enabled.
- *
- * In dev mode (via `npm run dev`), falls back to `import.meta.env.DEV`'s truthy
- * value since the artifact is not produced.
- */
-
-import { browser } from '$app/environment';
-import { base } from '$app/paths';
-
-let build = $state<string>('');
-
-async function loadBuild() {
-	if (!browser) return;
-
-	if (import.meta.env.DEV) {
-		build = 'dev';
-
-		return;
-	}
-
-	try {
-		const res = await fetch(`${base}/build.json`, { cache: 'no-store' });
-
-		if (res.ok) {
-			const data = await res.json();
-
-			build = data.version ?? '';
-		}
-	} catch {
-		// build.json missing or unreachable - leave as empty string
-	}
-}
-
-loadBuild();
-
-export const buildInfoStore = {
-	get value(): string {
-		return build;
-	}
-};

tools/ui/src/lib/stores/chat.svelte.ts

@@ -14,7 +14,6 @@
 import {
 	CONVERSATION_ID_SEPARATOR,
 	CWD_CLEARED_TEXT,
-	HEADERS,
 	INACTIVE_CONVERSATION,
 	STREAM_RESUME_RETRY_MS,
 	SYSTEM_MESSAGE_PLACEHOLDER,
@@ -25,7 +24,6 @@ import {
 	ErrorDialogType,
 	MessageRole,
 	MessageType,
-	MimeTypeApplication,
 	ReasoningEffort,
 	StreamConnectionState
 } from '$lib/enums';
@@ -58,7 +56,7 @@ import {
 	findMessageById,
 	formatCwdMessage,
 	generateConversationTitle,
-	getAuthHeaders,
+	getConversationModel,
 	isAbortError,
 	normalizeModelName,
 	streamIdentity
@@ -109,9 +107,6 @@ class ChatStore {
 	private isEditModeActive = $state(false);
 	private addFilesHandler: ((files: File[]) => void) | null = $state(null);
 	pendingEditMessageId = $state<string | null>(null);
-	private messageUpdateCallback:
-		| ((messageId: string, updates: Partial<DatabaseMessage>) => void)
-		| null = null;
 	private _pendingDraftMessage = $state<string>('');
 	private _pendingDraftFiles = $state<ChatUploadedFile[]>([]);
 
@@ -225,33 +220,12 @@ class ChatStore {
 	async probeServerStream(convId: string): Promise<ApiStreamSession | null> {
 		if (!convId) return null;
 
-		let listResp: Response;
-
-		try {
-			// POST the one conv id we are probing
-			listResp = await fetch(`./v1/streams/lookup`, {
-				body: JSON.stringify({ conversation_ids: [convId] }),
-				headers: { ...getAuthHeaders(), [HEADERS.CONTENT_TYPE]: MimeTypeApplication.JSON },
-				method: 'POST'
-			});
-		} catch (e) {
-			console.warn('probeServerStream fetch failed:', e);
-
-			return null;
-		}
-
-		if (!listResp.ok) {
-			console.warn(`probeServerStream got HTTP ${listResp.status} for conv ${convId}`);
-
-			return null;
-		}
-
 		let sessions: ApiStreamSession[];
 
 		try {
-			sessions = (await listResp.json()) as ApiStreamSession[];
+			sessions = await ChatService.lookupStreamSessions([convId]);
 		} catch (e) {
-			console.warn('probeServerStream JSON parse failed:', e);
+			console.warn(`probeServerStream failed for conv ${convId}:`, e);
 
 			return null;
 		}
@@ -296,18 +270,9 @@ class ChatStore {
 		let response: Response;
 
 		try {
-			response = await fetch(`./v1/stream?conv_id=${encodeURIComponent(id)}&from=0`, {
-				headers: getAuthHeaders()
-			});
+			response = await ChatService.fetchStreamReplay(id);
 		} catch (e) {
-			console.error('attachServerStream replay fetch failed:', e);
-			unlock();
-
-			return;
-		}
-
-		if (!response.ok) {
-			console.warn(`attachServerStream replay got HTTP ${response.status} for conv ${convId}`);
+			console.error(`attachServerStream replay failed for conv ${convId}:`, e);
 			unlock();
 
 			return;
@@ -807,21 +772,9 @@ class ChatStore {
 		let sessions: ApiStreamSession[];
 
 		try {
-			const resp = await fetch('./v1/streams/lookup', {
-				body: JSON.stringify({ conversation_ids: lookupIds }),
-				headers: { ...getAuthHeaders(), [HEADERS.CONTENT_TYPE]: MimeTypeApplication.JSON },
-				method: 'POST'
-			});
-
-			if (!resp.ok) return;
-
-			const body = (await resp.json()) as unknown;
-
-			if (!Array.isArray(body)) return;
-
-			sessions = body as ApiStreamSession[];
+			sessions = await ChatService.lookupStreamSessions(lookupIds);
 		} catch (e) {
-			console.warn('syncRemoteRunningStreams fetch failed:', e);
+			console.warn('syncRemoteRunningStreams lookup failed:', e);
 
 			return;
 		}
@@ -1335,7 +1288,7 @@ class ChatStore {
 		let effectiveModel: string | null | undefined = undefined;
 
 		if (serverStore.isRouterMode) {
-			const conversationModel = this.getConversationModel(allMessages);
+			const conversationModel = getConversationModel(allMessages);
 
 			effectiveModel = modelOverride || modelsStore.selectedModelName || conversationModel;
 		}
@@ -2788,16 +2741,6 @@ class ChatStore {
 		}
 	}
 
-	getConversationModel(messages: DatabaseMessage[]): string | null {
-		for (let i = messages.length - 1; i >= 0; i--) {
-			const message = messages[i];
-
-			if (message.role === MessageRole.ASSISTANT && message.model) return message.model;
-		}
-
-		return null;
-	}
-
 	private getApiOptions(): Record<string, unknown> {
 		const currentConfig = settingsStore.config;
 		const hasValue = (value: unknown): boolean =>

tools/ui/src/lib/stores/context-stats.svelte.ts

@@ -49,23 +49,8 @@ function deriveLiveStats(state: ApiProcessingState | null): LiveStats | null {
 }
 
 class ContextStatsStore {
-	// Resolve the model the stats report context for: explicit selection >
-	// last assistant model > single-model mode (mirrors useChatScreenActiveModel).
-	activeModelId = $derived.by(() => {
-		if (!serverStore.isRouterMode) {
-			return modelsStore.singleModelName;
-		}
-
-		const selectedId = modelsStore.selectedModelId;
-
-		if (selectedId) {
-			const model = modelsStore.models.find((m) => m.id === selectedId);
-
-			if (model) return model.model;
-		}
-
-		return chatStore.getConversationModel(conversationsStore.activeMessages as DatabaseMessage[]);
-	});
+	// The canonical resolution lives in modelsStore.activeModelId.
+	activeModelId = $derived(modelsStore.activeModelId);
 
 	isActiveModelLoaded = $derived(
 		this.activeModelId !== null &&

tools/ui/src/lib/stores/conversations.svelte.ts

@@ -20,21 +20,9 @@
 
 import { browser } from '$app/environment';
 import { goto } from '$app/navigation';
-import {
-	EXPORT_CONV,
-	NEWLINE,
-	REASONING_EFFORT_DEFAULT_LOCALSTORAGE_KEY,
-	ROUTES,
-	ZIP_MAGIC
-} from '$lib/constants';
-import {
-	FileExtensionText,
-	MessageRole,
-	MimeTypeApplication,
-	MimeTypeText,
-	ReasoningEffort,
-	SessionRecordType
-} from '$lib/enums';
+import { REASONING_EFFORT_DEFAULT_LOCALSTORAGE_KEY, ROUTES } from '$lib/constants';
+import { MessageRole, ReasoningEffort } from '$lib/enums';
+import { ConversationTransferService } from '$lib/services/conversation-transfer.service';
 import { DatabaseService } from '$lib/services/database.service';
 import { MigrationService } from '$lib/services/migration.service';
 import { RouterService } from '$lib/services/router.service';
@@ -43,7 +31,6 @@ import { mcpStore } from '$lib/stores/mcp.svelte';
 import { settingsStore } from '$lib/stores/settings.svelte';
 import type { McpServerOverride } from '$lib/types/database';
 import { filterByLeafNodeId, findLeafNode, generateConversationTitle } from '$lib/utils';
-import { strFromU8, strToU8, unzipSync, zipSync } from 'fflate';
 import { SvelteSet } from 'svelte/reactivity';
 import { toast } from 'svelte-sonner';
 
@@ -100,14 +87,6 @@ class ConversationsStore {
 		localStorage.setItem(REASONING_EFFORT_DEFAULT_LOCALSTORAGE_KEY, this.pendingReasoningEffort);
 	}
 
-	/**
-	 * Callback for updating message content in chatStore.
-	 * Registered by chatStore to enable cross-store updates without circular dependency.
-	 */
-	private messageUpdateCallback:
-		| ((messageId: string, updates: Partial<DatabaseMessage>) => void)
-		| null = null;
-
 	/** In-flight init run; shared by concurrent callers, reset on failure to allow retry */
 	private initPromise: Promise<void> | null = null;
 
@@ -143,23 +122,6 @@ class ConversationsStore {
 		return this.initPromise;
 	}
 
-	/**
-	 * Alias for init() for backward compatibility.
-	 */
-	async initialize(): Promise<void> {
-		return this.init();
-	}
-
-	/**
-	 * Register a callback for message updates from other stores.
-	 * Called by chatStore during initialization.
-	 */
-	registerMessageUpdateCallback(
-		callback: (messageId: string, updates: Partial<DatabaseMessage>) => void
-	): void {
-		this.messageUpdateCallback = callback;
-	}
-
 	/**
 	 *
 	 *
@@ -504,7 +466,7 @@ class ConversationsStore {
 				return;
 			}
 
-			this.downloadConversationsArchive(exported);
+			ConversationTransferService.downloadConversationsArchive(exported);
 
 			toast.success(
 				exported.length === 1
@@ -976,247 +938,6 @@ class ConversationsStore {
 	 *
 	 */
 
-	/**
-	 * Generates a sanitized filename for a conversation export
-	 * @param conversation - The conversation metadata
-	 * @param msgs - Optional array of messages belonging to the conversation
-	 * @returns The generated filename string
-	 */
-	generateConversationFilename(
-		conversation: { id?: string; name?: string },
-		msgs?: DatabaseMessage[]
-	): string {
-		const conversationName = (conversation.name ?? '').trim().toLowerCase();
-		const sanitizedName = conversationName
-			.replace(EXPORT_CONV.NON_ALPHANUMERIC_REGEX, EXPORT_CONV.NONALNUM_REPLACEMENT)
-			.replace(EXPORT_CONV.MULTIPLE_UNDERSCORE_REGEX, '_')
-			.substring(0, EXPORT_CONV.NAME_SUFFIX_MAX_LENGTH);
-		// If we have messages, use the timestamp of the newest message
-		const referenceDate = msgs?.length
-			? new Date(Math.max(...msgs.map((m) => m.timestamp)))
-			: new Date();
-		const iso = referenceDate.toISOString().slice(0, EXPORT_CONV.ISO_TIMESTAMP_SLICE);
-		const formattedDate = iso
-			.replace(EXPORT_CONV.ISO_DATE_TIME_SEPARATOR, EXPORT_CONV.ISO_DATE_TIME_SEPARATOR_REPLACEMENT)
-			.replaceAll(EXPORT_CONV.ISO_TIME_SEPARATOR, EXPORT_CONV.ISO_TIME_SEPARATOR_REPLACEMENT);
-		const trimmedConvId = conversation.id?.slice(0, EXPORT_CONV.ID_TRIM_LENGTH) ?? '';
-
-		return `${formattedDate}_conv_${trimmedConvId}_${sanitizedName}${FileExtensionText.JSONL}`;
-	}
-
-	/**
-	 * Serializes a session (a conversation with its messages) as JSONL.
-	 * The first line is the session header (a `SessionRecordType.SESSION` record
-	 * carrying the conversation properties); each subsequent line is a single message.
-	 * @param data - The exported conversation payload
-	 * @returns The JSONL string (one record per line)
-	 */
-	serializeSessionToJsonl(data: ExportedConversation): string {
-		const { conv, messages } = data;
-		const sessionLine = JSON.stringify({
-			harness: EXPORT_CONV.HARNESS,
-			type: SessionRecordType.SESSION,
-			...conv
-		});
-		const messageLines = messages.map((message: DatabaseMessage) => {
-			// `toolCalls` is stored as a JSON string; drop it when empty, otherwise parse it.
-			const { toolCalls, ...rest } = message;
-			const normalized = toolCalls ? { ...rest, toolCalls: JSON.parse(toolCalls) } : rest;
-
-			return JSON.stringify({ message: normalized, type: SessionRecordType.MESSAGE });
-		});
-
-		return [sessionLine, ...messageLines].join(NEWLINE);
-	}
-
-	/**
-	 * Parses the JSONL session format produced by {@link serializeSessionToJsonl}.
-	 * A `SessionRecordType.SESSION` line starts a new session; following
-	 * `SessionRecordType.MESSAGE` lines are appended to it. Supports multiple
-	 * sessions in a single file.
-	 * @param text - The JSONL file contents
-	 * @returns The parsed conversations with their messages
-	 */
-	parseSessionsJsonl(text: string): ExportedConversation[] {
-		const sessions: ExportedConversation[] = [];
-
-		let current: ExportedConversation | null = null;
-
-		for (const line of text.split(NEWLINE)) {
-			const trimmed = line.trim();
-
-			if (!trimmed) continue;
-
-			const record = JSON.parse(trimmed);
-
-			if (record.type === SessionRecordType.SESSION) {
-				// Drop the discriminator and harness marker; the rest is the conversation.
-				const conv = { ...record };
-
-				delete conv.type;
-				delete conv.harness;
-				current = { conv: conv as DatabaseConversation, messages: [] };
-				sessions.push(current);
-			} else if (record.type === SessionRecordType.MESSAGE) {
-				if (!current) {
-					throw new Error('Invalid JSONL: message record before any session record');
-				}
-
-				const message = record.message as DatabaseMessage;
-
-				// `toolCalls` is parsed to an array on export; the DB stores it as a string.
-				if (message.toolCalls !== undefined && typeof message.toolCalls !== 'string') {
-					message.toolCalls = JSON.stringify(message.toolCalls);
-				}
-
-				current.messages.push(message);
-			}
-			// Ignore unknown record types for forward compatibility.
-		}
-
-		return sessions;
-	}
-
-	/**
-	 * Reports whether the text is the JSONL session format, whose first non-empty
-	 * line is a `SessionRecordType.SESSION` record. A legacy JSON export starts
-	 * with an array or an object that has no such discriminator.
-	 * @param text - The file contents
-	 */
-	private isSessionsJsonl(text: string): boolean {
-		const trimmed = text.trimStart();
-		const lineEnd = trimmed.indexOf(NEWLINE);
-		const firstLine = lineEnd === -1 ? trimmed : trimmed.slice(0, lineEnd);
-
-		try {
-			return JSON.parse(firstLine).type === SessionRecordType.SESSION;
-		} catch {
-			// Not a standalone JSON record, so not the JSONL format.
-			return false;
-		}
-	}
-
-	/**
-	 * Parses an import file into conversations, accepting the current JSONL and
-	 * ZIP formats as well as the legacy JSON format. The format comes from the
-	 * contents, so an import works whatever the file is named.
-	 * @param file - The user-selected file
-	 * @returns The parsed conversations with their messages
-	 */
-	async parseImportFile(file: File): Promise<ExportedConversation[]> {
-		const bytes = new Uint8Array(await file.arrayBuffer());
-
-		if (ZIP_MAGIC.every((byte, index) => bytes[index] === byte)) {
-			const entries = unzipSync(bytes);
-			const sessions: ExportedConversation[] = [];
-
-			for (const [entryName, entryBytes] of Object.entries(entries)) {
-				if (!entryName.toLowerCase().endsWith(FileExtensionText.JSONL)) continue;
-
-				sessions.push(...this.parseSessionsJsonl(strFromU8(entryBytes)));
-			}
-
-			return sessions;
-		}
-
-		const text = strFromU8(bytes);
-
-		if (this.isSessionsJsonl(text)) {
-			return this.parseSessionsJsonl(text);
-		}
-
-		// Legacy JSON format: an array of conversations or a single conversation object.
-		const parsed = JSON.parse(text);
-
-		if (Array.isArray(parsed)) {
-			return parsed;
-		}
-
-		if (parsed && typeof parsed === 'object' && 'conv' in parsed && 'messages' in parsed) {
-			return [parsed];
-		}
-
-		throw new Error(
-			'Invalid file format: expected array of conversations or single conversation object'
-		);
-	}
-
-	/**
-	 * Triggers a browser download of the provided exported conversation data
-	 * @param data - The exported conversation payload (a single conversation with its messages)
-	 * @param filename - Filename; if omitted, a deterministic name is generated
-	 */
-	downloadConversationFile(data: ExportedConversation, filename?: string): void {
-		const { conv: conversation, messages: msgs } = data;
-
-		if (!conversation) {
-			console.error('Invalid data: missing conversation');
-
-			return;
-		}
-
-		const downloadFilename = filename ?? this.generateConversationFilename(conversation, msgs);
-		const jsonl = this.serializeSessionToJsonl(data);
-		const blob = new Blob([jsonl], { type: MimeTypeText.JSONL });
-
-		this.triggerDownload(blob, downloadFilename);
-	}
-
-	/**
-	 * Triggers a browser download of multiple conversations as a `.zip`, one
-	 * `.jsonl` file per conversation.
-	 * @param data - The conversations to export
-	 */
-	downloadConversationsArchive(data: ExportedConversation[]): void {
-		if (data.length === 0) {
-			console.error('Invalid data: no conversations to export');
-
-			return;
-		}
-
-		const usedNames = new SvelteSet<string>();
-		const files: Record<string, Uint8Array> = {};
-
-		for (const session of data) {
-			const baseName = this.generateConversationFilename(session.conv, session.messages);
-
-			// Disambiguate any duplicate filenames within the archive.
-			let entryName = baseName;
-			let suffix = 1;
-
-			while (usedNames.has(entryName)) {
-				entryName = baseName.replace(
-					new RegExp(`${FileExtensionText.JSONL}$`),
-					`_${suffix++}${FileExtensionText.JSONL}`
-				);
-			}
-			usedNames.add(entryName);
-
-			files[entryName] = strToU8(this.serializeSessionToJsonl(session));
-		}
-
-		const archiveName = `${new Date().toISOString().split(EXPORT_CONV.ISO_DATE_TIME_SEPARATOR)[0]}_conversations${FileExtensionText.ZIP}`;
-		const zipped = zipSync(files);
-		const blob = new Blob([zipped], { type: MimeTypeApplication.ZIP });
-
-		this.triggerDownload(blob, archiveName);
-	}
-
-	/**
-	 * Triggers a browser download of a blob under the given filename.
-	 */
-	private triggerDownload(blob: Blob, filename: string): void {
-		const url = URL.createObjectURL(blob);
-		const a = document.createElement('a');
-
-		a.href = url;
-		a.download = filename;
-		document.body.appendChild(a);
-		a.click();
-		document.body.removeChild(a);
-		URL.revokeObjectURL(url);
-	}
-
 	/**
 	 * Downloads a single conversation as a JSONL file, serializing the full message tree.
 	 * @param convId - The conversation ID to download
@@ -1231,7 +952,7 @@ class ConversationsStore {
 
 		const messages = await DatabaseService.getConversationMessages(convId);
 
-		this.downloadConversationFile({ conv: conversation, messages });
+		ConversationTransferService.downloadConversationFile({ conv: conversation, messages });
 	}
 
 	/**
@@ -1251,8 +972,3 @@ class ConversationsStore {
 }
 
 export const conversationsStore = new ConversationsStore();
-
-// Auto-initialize in browser
-if (browser) {
-	conversationsStore.init();
-}

tools/ui/src/lib/stores/device.svelte.ts

@@ -1,5 +1,17 @@
+/**
+ * deviceStore - Browser environment signals
+ *
+ * Device capabilities, OS theme and viewport in one class store:
+ * deviceStore.isMobile, deviceStore.isIOSDevice / isIOSSafari / isWKWebView /
+ * isStandalone, deviceStore.systemTheme.isDark.
+ *
+ * UA-derived flags are static for the session; isStandalone and systemTheme
+ * track live media query changes.
+ */
+
 import { browser } from '$app/environment';
-import { MEDIA_QUERIES } from '$lib/constants';
+import { DEFAULT_MOBILE_BREAKPOINT, MEDIA_QUERIES } from '$lib/constants';
+import { MediaQuery } from 'svelte/reactivity';
 
 /**
  * iOS UA token detection.
@@ -17,53 +29,60 @@ const UA_PATTERNS = {
 	WEBVIEW_IOS: /CriOS|FxiOS|EdgiOS|GSA/
 } as const;
 
-interface DeviceContext {
+class DeviceStore {
 	/** Any iOS/iPadOS device, regardless of which app or browser embeds the page. */
-	isIOSDevice: boolean;
+	readonly isIOSDevice: boolean = false;
 	/** The Safari browser app on iOS, excluding other iOS browsers and WKWebViews. */
-	isIOSSafari: boolean;
+	readonly isIOSSafari: boolean = false;
 	/** Any WKWebView context on iOS: in-app browsers, embedded web views, and the
 	 *  third-party iOS browsers (all of which share the WKWebView engine). */
-	isWKWebView: boolean;
+	readonly isWKWebView: boolean = false;
 	/** PWA standalone mode: the page was launched from the home screen icon. */
-	isStandalone: boolean;
+	isStandalone = $state(false);
+	/** OS color scheme preference; the user override lives in settingsStore. */
+	readonly systemTheme = $state({ isDark: false });
+
+	private mobile = new MediaQuery(`max-width: ${DEFAULT_MOBILE_BREAKPOINT - 1}px`);
+
+	get isMobile(): boolean {
+		return this.mobile.current;
+	}
+
+	constructor() {
+		if (!browser) return;
+
+		const ua = navigator.userAgent;
+		const isTouch = navigator.maxTouchPoints > 0;
+
+		this.isIOSDevice =
+			UA_PATTERNS.IOS_PHONE.test(ua) || (UA_PATTERNS.MACINTOSH.test(ua) && isTouch);
+		// Safari keeps 'Safari/' in the UA; non-Safari iOS browsers emit their own
+		// token instead. WKWebView typically omits 'Safari/' entirely.
+		const hasSafariToken = UA_PATTERNS.SAFARI.test(ua) && !UA_PATTERNS.WEBVIEW_IOS.test(ua);
+
+		this.isIOSSafari = this.isIOSDevice && hasSafariToken;
+		this.isWKWebView = this.isIOSDevice && !hasSafariToken;
+		// navigator.standalone is the legacy iOS-only flag (deprecated but still
+		// present); display-mode: standalone is the modern standard (Safari 16.4+).
+		this.isStandalone =
+			window.matchMedia(MEDIA_QUERIES.DISPLAY_MODE_STANDALONE).matches ||
+			(navigator as Navigator & { standalone?: boolean }).standalone === true;
+		this.systemTheme.isDark = window.matchMedia(MEDIA_QUERIES.PREFERS_DARK).matches;
+
+		// isStandalone and systemTheme can change at runtime (e.g. user installs the
+		// PWA while the tab is open); the UA-derived flags are static for the session
+		const standaloneMql = window.matchMedia(MEDIA_QUERIES.DISPLAY_MODE_STANDALONE);
+
+		standaloneMql.addEventListener('change', (e) => {
+			this.isStandalone = e.matches;
+		});
+
+		const darkMql = window.matchMedia(MEDIA_QUERIES.PREFERS_DARK);
+
+		darkMql.addEventListener('change', (e) => {
+			this.systemTheme.isDark = e.matches;
+		});
+	}
 }
 
-const SERVER_DEFAULT: DeviceContext = {
-	isIOSDevice: false,
-	isIOSSafari: false,
-	isStandalone: false,
-	isWKWebView: false
-};
-
-function detect(): DeviceContext {
-	if (!browser) return SERVER_DEFAULT;
-
-	const ua = navigator.userAgent;
-	const isTouch = navigator.maxTouchPoints > 0;
-	const isIOSDevice = UA_PATTERNS.IOS_PHONE.test(ua) || (UA_PATTERNS.MACINTOSH.test(ua) && isTouch);
-	// Safari keeps 'Safari/' in the UA; non-Safari iOS browsers emit their own
-	// token instead. WKWebView typically omits 'Safari/' entirely.
-	const hasSafariToken = UA_PATTERNS.SAFARI.test(ua) && !UA_PATTERNS.WEBVIEW_IOS.test(ua);
-	const isIOSSafari = isIOSDevice && hasSafariToken;
-	const isWKWebView = isIOSDevice && !hasSafariToken;
-	// navigator.standalone is the legacy iOS-only flag (deprecated but still
-	// present); display-mode: standalone is the modern standard (Safari 16.4+).
-	const isStandalone =
-		window.matchMedia(MEDIA_QUERIES.DISPLAY_MODE_STANDALONE).matches ||
-		(navigator as Navigator & { standalone?: boolean }).standalone === true;
-
-	return { isIOSDevice, isIOSSafari, isStandalone, isWKWebView };
-}
-
-export const device = $state<DeviceContext>(detect());
-
-if (browser) {
-	// isStandalone can change at runtime (e.g. user installs the PWA while the
-	// tab is open); the UA-derived flags are static for the session.
-	const mql = window.matchMedia(MEDIA_QUERIES.DISPLAY_MODE_STANDALONE);
-
-	mql.addEventListener('change', (e) => {
-		device.isStandalone = e.matches;
-	});
-}
+export const deviceStore = new DeviceStore();

tools/ui/src/lib/stores/index.ts

@@ -53,26 +53,6 @@ export { permissionsStore } from './permissions.svelte';
 export { toolsStore } from './tools.svelte';
 
 // ENVIRONMENT / META
-export { buildInfoStore } from './build-info.svelte';
-
 export { versionStore } from './version.svelte';
 
-export { device } from './device.svelte';
-
-export { viewport, isMobile } from './viewport.svelte';
-
-export { theme } from './theme.svelte';
-
-export {
-	gaugePopup,
-	gaugePopupClose,
-	gaugeTriggerPointerDown,
-	gaugeTriggerClick,
-	gaugeTriggerKeydown,
-	gaugeTriggerEnter,
-	gaugeTriggerLeave,
-	gaugeCardEnter,
-	gaugeCardLeave
-} from './context-gauge-popup.svelte';
-
-export { persisted } from './persisted.svelte';
+export { deviceStore } from './device.svelte';

tools/ui/src/lib/stores/init.ts

@@ -0,0 +1,34 @@
+/**
+ * Explicit store initialization, called once from the root layout.
+ *
+ * Order matters: migrations run first because they rename and rewrite
+ * localStorage keys, so every store that reads localStorage initializes
+ * only after they complete. Constructors and module-level side effects
+ * stay empty so import order can no longer change startup behavior.
+ */
+
+// direct imports, not via the barrel, to avoid circular deps
+import { conversationsStore } from './conversations.svelte';
+import { permissionsStore } from './permissions.svelte';
+import { settingsStore } from './settings.svelte';
+import { toolsStore } from './tools.svelte';
+import { versionStore } from './version.svelte';
+import { browser } from '$app/environment';
+import { MigrationService } from '$lib/services/migration.service';
+
+let started = false;
+
+export async function initStores(): Promise<void> {
+	if (!browser || started) return;
+
+	started = true;
+
+	await MigrationService.runAllMigrations();
+
+	settingsStore.initialize();
+	permissionsStore.initialize();
+	toolsStore.initialize();
+	void versionStore.initialize();
+
+	await conversationsStore.init();
+}

tools/ui/src/lib/stores/models.svelte.ts

@@ -1,12 +1,4 @@
-import { base } from '$app/paths';
-import {
-	API_MODELS,
-	FAVORITE_MODELS_LOCALSTORAGE_KEY,
-	MODEL_PROPS_CACHE,
-	SSE_DATA_PREFIX,
-	SSE_LINE_SEPARATOR,
-	SSE_RECORD_SEPARATOR
-} from '$lib/constants';
+import { FAVORITE_MODELS_LOCALSTORAGE_KEY, MODEL_PROPS_CACHE } from '$lib/constants';
 import {
 	FileTypeCategory,
 	ModelModality,
@@ -20,12 +12,12 @@ import { conversationsStore } from '$lib/stores/conversations.svelte';
 import { serverStore } from '$lib/stores/server.svelte';
 // deep imports, not the '$lib/utils' barrel: it re-exports modules that reach back
 // into the stores, and going through it here would read a half-built module
-import { getAuthHeaders } from '$lib/utils/api-headers';
 import { TTLCache } from '$lib/utils/cache-ttl';
 import {
 	detectThinkingSupport,
 	detectThinkingSupportWithReason
 } from '$lib/utils/chat-template-thinking-detector';
+import { getConversationModel } from '$lib/utils/conversation-utils';
 import { SvelteMap, SvelteSet } from 'svelte/reactivity';
 import { toast } from 'svelte-sonner';
 
@@ -141,6 +133,33 @@ class ModelsStore {
 		return props.model_path.split(/(\\|\/)/).pop() || null;
 	}
 
+	/**
+	 * Model the active conversation view resolves to. Router mode: the user's
+	 * selection first, then the conversation's own model. Otherwise the single
+	 * served model, from the models list or the server props as a fallback.
+	 */
+	get activeModelId(): string | null {
+		if (!serverStore.isRouterMode) {
+			return this.models.length > 0 ? this.models[0].model : this.singleModelName;
+		}
+
+		if (this.selectedModelId) {
+			const selected = this.models.find((m) => m.id === this.selectedModelId);
+
+			if (selected) return selected.model;
+		}
+
+		const conversationModel = getConversationModel(conversationsStore.activeMessages);
+
+		if (conversationModel) {
+			const model = this.models.find((m) => m.model === conversationModel);
+
+			if (model) return model.model;
+		}
+
+		return null;
+	}
+
 	get selectedModelContextSize(): number | null {
 		if (!this.selectedModelName) return null;
 
@@ -717,8 +736,6 @@ class ModelsStore {
 	 */
 
 	// reconnect delay after the feed drops or the server is not ready yet
-	private static readonly SSE_RECONNECT_MS = 1000;
-
 	/**
 	 * Open the /models/sse feed and keep it live with auto reconnect.
 	 * Idempotent and router mode only. The feed drives status and progress,
@@ -752,72 +769,10 @@ class ModelsStore {
 	}
 
 	/**
-	 * Read the feed and reconnect until unsubscribed. Splits the byte stream
-	 * into SSE records on the blank line boundary.
+	 * Read the feed and reconnect until unsubscribed.
 	 */
 	private async runStatusReader(signal: AbortSignal): Promise<void> {
-		const decoder = new TextDecoder();
-
-		while (!signal.aborted) {
-			try {
-				const response = await fetch(`${base}${API_MODELS.SSE}`, {
-					headers: getAuthHeaders(),
-					signal
-				});
-
-				if (response.ok && response.body) {
-					const reader = response.body.getReader();
-
-					let buffer = '';
-
-					while (!signal.aborted) {
-						const { done, value } = await reader.read();
-
-						if (done) break;
-
-						buffer += decoder.decode(value, { stream: true });
-
-						let boundary = buffer.indexOf(SSE_RECORD_SEPARATOR);
-
-						while (boundary !== -1) {
-							this.handleStatusRecord(buffer.slice(0, boundary));
-							buffer = buffer.slice(boundary + SSE_RECORD_SEPARATOR.length);
-							boundary = buffer.indexOf(SSE_RECORD_SEPARATOR);
-						}
-					}
-				}
-			} catch {
-				// network drop or abort falls through to the reconnect delay
-			}
-
-			if (signal.aborted) return;
-
-			await new Promise((resolve) => setTimeout(resolve, ModelsStore.SSE_RECONNECT_MS));
-		}
-	}
-
-	/**
-	 * Parse one SSE record. The payload rides in the data lines as a JSON
-	 * envelope that carries its own model, event and data fields.
-	 */
-	private handleStatusRecord(record: string): void {
-		const payload = record
-			.split(SSE_LINE_SEPARATOR)
-			.filter((line) => line.startsWith(SSE_DATA_PREFIX))
-			.map((line) => line.slice(SSE_DATA_PREFIX.length).trim())
-			.join(SSE_LINE_SEPARATOR);
-
-		if (payload.length === 0) return;
-
-		let envelope: ApiModelsSseEvent;
-
-		try {
-			envelope = JSON.parse(payload);
-		} catch {
-			return;
-		}
-
-		this.applyStatusEvent(envelope);
+		await ModelsService.watchModelEvents(signal, (event) => this.applyStatusEvent(event));
 	}
 
 	/**

tools/ui/src/lib/stores/permissions.svelte.ts

@@ -5,7 +5,11 @@ import { SvelteSet } from 'svelte/reactivity';
 class PermissionsStore {
 	private _tools = $state(new SvelteSet<string>());
 
-	constructor() {
+	/**
+	 * Load persisted permissions. Called by initStores() after migrations
+	 * have run.
+	 */
+	initialize(): void {
 		// browser-only init: skip on SSR to avoid localStorage side effects
 		if (!browser) return;
 

tools/ui/src/lib/stores/persisted.svelte.ts

@@ -1,51 +0,0 @@
-import { browser } from '$app/environment';
-
-type PersistedValue<T> = {
-	get value(): T;
-	set value(newValue: T);
-};
-
-export function persisted<T>(key: string, initialValue: T): PersistedValue<T> {
-	let value = initialValue;
-
-	if (browser) {
-		try {
-			const stored = localStorage.getItem(key);
-
-			if (stored !== null) {
-				value = JSON.parse(stored) as T;
-			}
-		} catch (error) {
-			console.warn(`Failed to load ${key}:`, error);
-		}
-	}
-
-	const persist = (next: T) => {
-		if (!browser) {
-			return;
-		}
-
-		try {
-			if (next === null || next === undefined) {
-				localStorage.removeItem(key);
-
-				return;
-			}
-
-			localStorage.setItem(key, JSON.stringify(next));
-		} catch (error) {
-			console.warn(`Failed to persist ${key}:`, error);
-		}
-	};
-
-	return {
-		get value() {
-			return value;
-		},
-
-		set value(newValue: T) {
-			value = newValue;
-			persist(newValue);
-		}
-	};
-}

tools/ui/src/lib/stores/settings.svelte.ts

@@ -40,9 +40,9 @@ import {
 } from '$lib/constants';
 import { ColorMode } from '$lib/enums';
 import { ParameterSyncService } from '$lib/services/parameter-sync.service';
+import { deviceStore } from '$lib/stores/device.svelte';
 // direct imports between stores, not via the barrel, to avoid circular deps
 import { serverStore } from '$lib/stores/server.svelte';
-import { isMobile } from '$lib/stores/viewport.svelte';
 import type { SettingsExportType } from '$lib/types';
 import {
 	configToParameterRecord,
@@ -85,12 +85,6 @@ class SettingsStore {
 		return ParameterSyncService.extractServerDefaults(serverStore.defaultParams);
 	}
 
-	constructor() {
-		if (browser) {
-			this.initialize();
-		}
-	}
-
 	/**
 	 *
 	 *
@@ -100,9 +94,12 @@ class SettingsStore {
 	 */
 
 	/**
-	 * Initialize the settings store by loading from localStorage
+	 * Initialize the settings store by loading from localStorage.
+	 * Called by initStores() after migrations have run.
 	 */
 	initialize() {
+		if (!browser) return;
+
 		try {
 			this.loadConfig();
 			this.migrateLegacyTheme();
@@ -138,7 +135,7 @@ class SettingsStore {
 
 			// Default sendOnEnter to false on mobile when the user has no saved preference
 			if (!(SETTINGS_KEYS.SEND_ON_ENTER in savedVal)) {
-				if (isMobile.current) {
+				if (deviceStore.isMobile) {
 					this.config[SETTINGS_KEYS.SEND_ON_ENTER] = false;
 				}
 			}

tools/ui/src/lib/stores/theme.svelte.ts

@@ -1,14 +0,0 @@
-import { browser } from '$app/environment';
-import { MEDIA_QUERIES } from '$lib/constants';
-
-export const theme = $state({
-	isSystemDark: browser && window.matchMedia(MEDIA_QUERIES.PREFERS_DARK).matches
-});
-
-if (browser) {
-	const mql = window.matchMedia(MEDIA_QUERIES.PREFERS_DARK);
-
-	mql.addEventListener('change', (e) => {
-		theme.isSystemDark = e.matches;
-	});
-}

tools/ui/src/lib/stores/tools.svelte.ts

@@ -38,7 +38,11 @@ class ToolsStore {
 	private _toolsEndpointUnreachable = $state(false);
 	private _serverHome = $state<string | null | undefined>(undefined);
 
-	constructor() {
+	/**
+	 * Load persisted disabled tools and fetch the builtin tool list.
+	 * Called by initStores() after migrations have run.
+	 */
+	initialize(): void {
 		// browser-only init: skip on SSR to avoid localStorage/fetch side effects
 		if (!browser) return;
 

tools/ui/src/lib/stores/version.svelte.ts

@@ -1,44 +1,65 @@
 /**
- * versionStore - Frontend build version
+ * versionStore - Build version information
  *
- * Reads from SvelteKit's `_app/version.json` — generated by the @vite-pwa/sveltekit
- * plugin. The version string changes on every build, so comparing it against
- * localStorage reliably detects server upgrades.
+ * - `build`: llama.cpp build number from `build.json`, embedded at llama.cpp
+ *   build time (LLAMA_BUILD_NUMBER). Shown in the UI when `showBuildVersion`
+ *   is enabled.
+ * - `frontend`: frontend build version from SvelteKit's `_app/version.json`,
+ *   generated by the @vite-pwa/sveltekit plugin. Changes on every build, so
+ *   comparing it against localStorage reliably detects server upgrades.
  *
- * In dev mode, falls back to `'dev'`.
+ * In dev mode both fall back to `'dev'`.
  */
 
 import { browser } from '$app/environment';
 import { base } from '$app/paths';
 
-let version = $state<string>('');
+class VersionStore {
+	build = $state<string>('');
+	frontend = $state<string>('');
 
-async function loadVersion() {
-	if (!browser) return;
+	/**
+	 * Fetch the version files. Called by initStores(); order-independent,
+	 * so it runs in the background.
+	 */
+	initialize(): void {
+		if (!browser) return;
 
-	if (import.meta.env.DEV) {
-		version = 'dev';
+		if (import.meta.env.DEV) {
+			this.build = 'dev';
+			this.frontend = 'dev';
 
-		return;
+			return;
+		}
+
+		void this.load();
 	}
 
-	try {
-		const res = await fetch(`${base}/_app/version.json`, { cache: 'no-store' });
+	private async load(): Promise<void> {
+		try {
+			const res = await fetch(`${base}/build.json`, { cache: 'no-store' });
 
-		if (res.ok) {
-			const data = await res.json();
+			if (res.ok) {
+				const data = await res.json();
 
-			version = data.version ?? '';
+				this.build = data.version ?? '';
+			}
+		} catch {
+			// build.json missing or unreachable - leave as empty string
+		}
+
+		try {
+			const res = await fetch(`${base}/_app/version.json`, { cache: 'no-store' });
+
+			if (res.ok) {
+				const data = await res.json();
+
+				this.frontend = data.version ?? '';
+			}
+		} catch {
+			// version.json missing or unreachable - leave as empty string
 		}
-	} catch {
-		// _app/version.json missing or unreachable - leave as empty string
 	}
 }
 
-loadVersion();
-
-export const versionStore = {
-	get value(): string {
-		return version;
-	}
-};
+export const versionStore = new VersionStore();

tools/ui/src/lib/stores/viewport.svelte.ts

@@ -1,9 +0,0 @@
-import { browser } from '$app/environment';
-import { DEFAULT_MOBILE_BREAKPOINT } from '$lib/constants';
-import { MediaQuery } from 'svelte/reactivity';
-
-export const viewport = $state({
-	width: browser ? window.innerWidth : 0
-});
-
-export const isMobile = new MediaQuery(`max-width: ${DEFAULT_MOBILE_BREAKPOINT - 1}px`);

tools/ui/src/lib/utils/conversation-utils.ts

@@ -1,8 +1,23 @@
 /**
  * Utility functions for conversation data manipulation
  */
+import { MessageRole } from '$lib/enums';
 import type { DatabaseMessage } from '$lib/types';
 
+/**
+ * Model that generated the conversation's latest assistant message, or null
+ * when no assistant message carries one.
+ */
+export function getConversationModel(messages: readonly DatabaseMessage[]): string | null {
+	for (let i = messages.length - 1; i >= 0; i--) {
+		const message = messages[i];
+
+		if (message.role === MessageRole.ASSISTANT && message.model) return message.model;
+	}
+
+	return null;
+}
+
 /**
  * Creates a map of conversation IDs to their message counts from exported conversation data
  * @param exportedData - Array of exported conversations with their messages

tools/ui/src/lib/utils/index.ts

@@ -54,6 +54,7 @@ export { modelLoadFraction, modelLoadProgressText } from './progress';
 export {
 	createMessageCountMap,
 	getMessageCount,
+	getConversationModel,
 	buildConversationTree,
 	type ConversationTreeItem
 } from './conversation-utils';

tools/ui/src/routes/(chat)/+page.svelte

@@ -77,7 +77,7 @@
 
 	onMount(async () => {
 		if (!conversationsStore.isInitialized) {
-			await conversationsStore.initialize();
+			await conversationsStore.init();
 		}
 
 		conversationsStore.clearActiveConversation();

tools/ui/src/routes/+layout.svelte

@@ -19,16 +19,16 @@
 	import { usePwa } from '$lib/hooks/use-pwa.svelte';
 	import { RouterService } from '$lib/services/router.service';
 	import {
-		buildInfoStore,
 		chatStore,
 		conversationsStore,
-		isMobile,
+		deviceStore,
 		mcpStore,
 		modelsStore,
 		serverStore,
 		settingsStore,
-		theme
+		versionStore
 	} from '$lib/stores';
+	import { initStores } from '$lib/stores/init';
 	import { ModeWatcher } from 'mode-watcher';
 	import { untrack } from 'svelte';
 	import { onMount } from 'svelte';
@@ -36,6 +36,10 @@
 	import { pwaAssetsHead } from 'virtual:pwa-assets/head';
 
 	let { children } = $props();
+
+	// migrations and store startup, ordered explicitly instead of import side effects
+	void initStores();
+
 	let innerHeight = $state<number | undefined>();
 	let innerWidth = $state(browser ? window.innerWidth : 0);
 
@@ -55,7 +59,7 @@
 	const { needRefresh, updateServiceWorker } = pwa;
 
 	function updateFavicon() {
-		const dark = theme.isSystemDark;
+		const dark = deviceStore.systemTheme.isDark;
 
 		let icoLink = document.querySelector(FAVICON_SELECTORS.ICO_48X48) as HTMLLinkElement | null;
 
@@ -153,7 +157,7 @@
 	}
 
 	$effect(() => {
-		void theme.isSystemDark;
+		void deviceStore.systemTheme.isDark;
 
 		updateFavicon();
 	});
@@ -274,7 +278,7 @@
 	<div class="flex flex-col md:flex-row">
 		<SidebarNavigation
 			onSearchClick={() => {
-				if (isMobile.current) {
+				if (deviceStore.isMobile) {
 					goto(ROUTES.SEARCH);
 				} else if (chatSidebar?.activateSearchMode) {
 					chatSidebar.activateSearchMode();
@@ -294,8 +298,8 @@
 
 <!-- PWA update prompt + version -->
 <div class="fixed right-4 bottom-4 z-9999 flex flex-col items-end gap-1">
-	{#if showBuildVersion && buildInfoStore.value}
-		<span class="text-[10px] tabular-nums text-muted-foreground">{buildInfoStore.value}</span>
+	{#if showBuildVersion && versionStore.build}
+		<span class="text-[10px] tabular-nums text-muted-foreground">{versionStore.build}</span>
 	{/if}
 
 	<PwaRefreshAlert

tools/ui/src/routes/search/+page.svelte

@@ -5,7 +5,7 @@
 	import { SearchInput, SidebarNavigationSearchResults } from '$lib/components/app';
 	import { ROUTES } from '$lib/constants';
 	import { RouterService } from '$lib/services/router.service';
-	import { chatStore, conversationsStore, isMobile } from '$lib/stores';
+	import { chatStore, conversationsStore, deviceStore } from '$lib/stores';
 
 	let searchQuery = $state('');
 	let searchInputRef = $state<HTMLInputElement | null>(null);
@@ -23,7 +23,7 @@
 	// Search page is intended for mobile; on desktop the sidebar already exposes
 	// in-place search, so bounce back to a chat.
 	$effect(() => {
-		if (browser && !isMobile.current) {
+		if (browser && !deviceStore.isMobile) {
 			goto(ROUTES.NEW_CHAT, { replaceState: true });
 		}
 	});

tools/ui/tests/unit/conversation-import.test.ts

@@ -1,35 +1,9 @@
 import { NEWLINE } from '$lib/constants';
 import { MessageRole, MessageType } from '$lib/enums';
+import { ConversationTransferService } from '$lib/services/conversation-transfer.service';
 import type { ExportedConversation } from '$lib/types/database';
 import { strToU8, zipSync } from 'fflate';
-import { beforeAll, describe, expect, it } from 'vitest';
-
-let conversationsStore: typeof import('$lib/stores/conversations.svelte').conversationsStore;
-
-// node env unit project has no DOM, install a minimal localStorage backed by a
-// Map before the store module reads it. Transforming the store takes seconds,
-// so import it once for the whole file.
-beforeAll(async () => {
-	const store = new Map<string, string>();
-	const polyfill: Storage = {
-		clear: () => store.clear(),
-		getItem: (k) => (store.has(k) ? store.get(k)! : null),
-		key: (i) => Array.from(store.keys())[i] ?? null,
-		get length() {
-			return store.size;
-		},
-		removeItem: (k) => {
-			store.delete(k);
-		},
-		setItem: (k, v) => {
-			store.set(k, String(v));
-		}
-	};
-
-	(globalThis as unknown as { localStorage: Storage }).localStorage = polyfill;
-
-	({ conversationsStore } = await import('$lib/stores/conversations.svelte'));
-}, 30000);
+import { describe, expect, it } from 'vitest';
 
 function makeSession(id: string): ExportedConversation {
 	return {
@@ -54,10 +28,10 @@ function makeSession(id: string): ExportedConversation {
  * for `.jsonl`, so the picker cannot filter on it and the filename carries no
  * guarantee: a JSONL export must import under any name.
  */
-describe('conversationsStore.parseImportFile', () => {
+describe('ConversationTransferService.parseImportFile', () => {
 	it('imports a JSONL export whose name has no meaningful extension', async () => {
-		const jsonl = conversationsStore.serializeSessionToJsonl(makeSession('a'));
-		const sessions = await conversationsStore.parseImportFile(new File([jsonl], 'export'));
+		const jsonl = ConversationTransferService.serializeSessionToJsonl(makeSession('a'));
+		const sessions = await ConversationTransferService.parseImportFile(new File([jsonl], 'export'));
 
 		expect(sessions).toHaveLength(1);
 		expect(sessions[0].conv.id).toBe('a');
@@ -66,27 +40,33 @@ describe('conversationsStore.parseImportFile', () => {
 
 	it('imports several sessions from one JSONL file', async () => {
 		const jsonl = [makeSession('a'), makeSession('b')]
-			.map((session) => conversationsStore.serializeSessionToJsonl(session))
+			.map((session) => ConversationTransferService.serializeSessionToJsonl(session))
 			.join(NEWLINE);
-		const sessions = await conversationsStore.parseImportFile(new File([jsonl], 'export.txt'));
+		const sessions = await ConversationTransferService.parseImportFile(
+			new File([jsonl], 'export.txt')
+		);
 
 		expect(sessions.map((session) => session.conv.id)).toEqual(['a', 'b']);
 	});
 
 	it('imports a ZIP archive whose name has no meaningful extension', async () => {
 		const zipped = zipSync({
-			'a.jsonl': strToU8(conversationsStore.serializeSessionToJsonl(makeSession('a'))),
-			'b.jsonl': strToU8(conversationsStore.serializeSessionToJsonl(makeSession('b'))),
+			'a.jsonl': strToU8(ConversationTransferService.serializeSessionToJsonl(makeSession('a'))),
+			'b.jsonl': strToU8(ConversationTransferService.serializeSessionToJsonl(makeSession('b'))),
 			'notes.txt': strToU8('ignored')
 		});
-		const sessions = await conversationsStore.parseImportFile(new File([zipped], 'archive'));
+		const sessions = await ConversationTransferService.parseImportFile(
+			new File([zipped], 'archive')
+		);
 
 		expect(sessions.map((session) => session.conv.id).sort()).toEqual(['a', 'b']);
 	});
 
 	it('imports the legacy JSON array format', async () => {
 		const json = JSON.stringify([makeSession('a')], null, 2);
-		const sessions = await conversationsStore.parseImportFile(new File([json], 'export.jsonl'));
+		const sessions = await ConversationTransferService.parseImportFile(
+			new File([json], 'export.jsonl')
+		);
 
 		expect(sessions).toHaveLength(1);
 		expect(sessions[0].conv.id).toBe('a');
@@ -94,7 +74,7 @@ describe('conversationsStore.parseImportFile', () => {
 
 	it('imports the legacy JSON single object format', async () => {
 		const json = JSON.stringify(makeSession('a'));
-		const sessions = await conversationsStore.parseImportFile(new File([json], 'export'));
+		const sessions = await ConversationTransferService.parseImportFile(new File([json], 'export'));
 
 		expect(sessions).toHaveLength(1);
 		expect(sessions[0].conv.id).toBe('a');
@@ -102,7 +82,7 @@ describe('conversationsStore.parseImportFile', () => {
 
 	it('rejects a file that holds neither format', async () => {
 		await expect(
-			conversationsStore.parseImportFile(new File(['not an export'], 'export.jsonl'))
+			ConversationTransferService.parseImportFile(new File(['not an export'], 'export.jsonl'))
 		).rejects.toThrow();
 	});
 });