ci : run ccache-clear as the last step of release jobs (#27503 )

* ci : run ccache-clear as the last step of release jobs Assisted-by: pi:llama.cpp/Qwen3.8-27B * update disabled job too to force rebase --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
llama.cpp : bump version to 0.2.0 (#27498 )
2026-08-21 23:57:44 +02:00 · 2026-08-21 18:56:04 +03:00 · 2026-08-21 15:01:24 +03:00 · 2026-08-21 14:51:26 +03:00 · 2026-08-21 14:36:36 +03:00 · 2026-08-21 14:36:36 +03:00
163 changed files with 12692 additions and 13668 deletions
@@ -1,3 +1,4 @@
+# note: place this as the last step of the job, so the new cache is saved by "Post ccache" right after the old one is cleared
 name: "ccache-clear"
 description: "Delete all GitHub Actions caches matching a key prefix"
 inputs:
@@ -27,30 +27,26 @@ jobs:
          cmake --install build --prefix "$PREFIX" --config Release

          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
-          tclsh <<'EOF'
-          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
-          set build(number)  [string trim [exec git rev-list  --count HEAD]]
+          build_commit=$(git rev-parse --short HEAD | xargs)
+          build_number=$(git rev-list --count HEAD | xargs)

-          set cmakelists [read [open "CMakeLists.txt" r]]
-          regexp {set\(LLAMA_VERSION_MAJOR\s+(\d+)\)} $cmakelists -> major
-          regexp {set\(LLAMA_VERSION_MINOR\s+(\d+)\)} $cmakelists -> minor
-          regexp {set\(LLAMA_VERSION_PATCH\s+(\d+)\)} $cmakelists -> patch
-          set build(version) "$major.$minor.$patch"
+          major=$(grep -oE "set\(LLAMA_VERSION_MAJOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          minor=$(grep -oE "set\(LLAMA_VERSION_MINOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          patch=$(grep -oE "set\(LLAMA_VERSION_PATCH[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          build_version="$major.$minor.$patch"

-          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
-          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
-                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
-                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
+          checks=("set\(LLAMA_VERSION[[:space:]]+$build_version\)"
+                  "set\(LLAMA_BUILD_COMMIT[[:space:]]+$build_commit\)"
+                  "set\(LLAMA_BUILD_NUMBER[[:space:]]+$build_number\)")

-          puts -nonewline "Checking llama-config.cmake version... "
-          foreach check $checks {
-              if {![regexp -expanded -- $check $llamaconfig]} {
-                  puts "\"$check\" failed!"
+          for check in "${checks[@]}"; do
+              if ! grep -qE "$check" "$LLAMA_CONFIG"; then
+                  echo "Checking llama-config.cmake version... \"$check\" failed!"
                  exit 1
-              }
-          }
-          puts "success."
-          EOF
+              fi
+          done
+
+          echo "Checking llama-config.cmake version... success."

          cd examples/simple-cmake-pkg
          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
@@ -97,8 +97,7 @@ jobs:
          cmake -B build \
            -DGGML_NATIVE=OFF \
            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_RPC=ON \
-            -DGGML_NATIVE=OFF
+            -DGGML_RPC=ON
          time cmake --build build --config Release -j $(nproc)

      - name: Test
@@ -55,6 +55,20 @@ jobs:
        env:
          GITHUB_REPOSITORY: ${{ github.repository }}

+      - name: Create nightly-tag.txt
+        id: nightly_tag_file
+        run: |
+          NIGHTLY_TAG="${{ steps.desc.outputs.nightly_tag }}"
+          if [[ -z "${NIGHTLY_TAG}" ]]; then
+            echo "Warning: no nightly tag found for the release commit - nightly-tag.txt will not be created"
+            echo "create=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+          echo "${NIGHTLY_TAG}" > nightly-tag.txt
+          echo "create=true" >> "$GITHUB_OUTPUT"
+          echo "nightly-tag.txt:"
+          cat nightly-tag.txt
+
      - name: Create release
        if: ${{ github.event.inputs.dry_run == 'false' }}
        uses: ggml-org/action-create-release@v1
@@ -62,26 +76,49 @@ jobs:
          GITHUB_TOKEN: ${{ github.token }}
        with:
          tag_name: ${{ steps.checks.outputs.version }}
-          # TODO: remove the prerelease flag once the semantic versioning workflow is ready
-          # ref: https://github.com/ggml-org/ggml/discussions/1579
-          prerelease: true
+          prerelease: false
+          # TODO: enrich the body of the release with more information
          body: |
-            > [!NOTE]
-            > Semantic versioning is still work in progress.
-            > More info can be found in https://github.com/ggml-org/ggml/discussions/1579
+            ## Overview
+
+            New version has been released.

            ${{ steps.desc.outputs.nightly }}

+            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
+
+            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
+
            ## ${{ steps.desc.outputs.changelog_title }}

            ${{ steps.desc.outputs.changelog }}

+      - name: Upload nightly-tag.txt
+        if: ${{ github.event.inputs.dry_run == 'false' && steps.nightly_tag_file.outputs.create == 'true' }}
+        uses: actions/github-script@v8
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            console.log('uploadReleaseAsset', 'nightly-tag.txt');
+            await github.rest.repos.uploadReleaseAsset({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              release_id: release_id,
+              name: 'nightly-tag.txt',
+              data: await fs.readFileSync('./nightly-tag.txt')
+            });
+
      - name: Dry run summary
        if: ${{ github.event.inputs.dry_run == 'true' }}
        run: |
          if [[ "${{ steps.checks.outputs.checks_passed }}" == "true" ]]; then
            echo "Dry run complete - all checks passed."
            echo "Would have created tag: ${{ steps.checks.outputs.version }}"
+            if [[ -n "${{ steps.desc.outputs.nightly_tag }}" ]]; then
+              echo "Would have uploaded nightly-tag.txt: ${{ steps.desc.outputs.nightly_tag }}"
+            fi
          else
            echo "::error::Dry run found release check failures. A release tag would not be created."
            exit 1
@@ -145,11 +145,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-${{ matrix.arch }}
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -166,6 +161,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-macos-${{ matrix.build }}.tar.gz
          name: llama-bin-macos-${{ matrix.build }}.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-${{ matrix.arch }}
+
  ubuntu-cpu:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -231,12 +231,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        if: ${{ matrix.build != 's390x' }}
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-cpu
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -253,6 +247,12 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
          name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz

+      - name: ccache-clear
+        if: ${{ matrix.build != 's390x' }}
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-cpu
+
  ubuntu-vulkan:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -318,11 +318,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-vulkan
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -339,6 +334,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
          name: llama-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-vulkan
+
  android-arm64:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -512,11 +512,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build/ReleaseOV --config Release --parallel

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-openvino-release-no-preset-v1
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -551,6 +546,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
          name: llama-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-openvino-release-no-preset-v1
+
  windows-openvino:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -637,11 +637,6 @@ jobs:

          cmake --build build\ReleaseOV --config Release -- /m

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-openvino
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -680,6 +675,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip
          name: llama-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-openvino
+
  windows-cpu:
    name: windows-cpu / ${{ matrix.arch }}
    needs: [check-release]
@@ -733,11 +733,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
-
      - name: Pack artifacts
        id: pack_artifacts
        run: |
@@ -749,6 +744,11 @@ jobs:
          path: llama-bin-win-cpu-${{ matrix.arch }}.zip
          name: llama-bin-win-cpu-${{ matrix.arch }}.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
+
  windows-rocm:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -841,11 +841,6 @@ jobs:
            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}"
          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-
      - name: Verify HIP backend was built
        run: |
          $hipDll = Get-ChildItem -Path build\bin -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
@@ -878,6 +873,11 @@ jobs:
          path: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
          name: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+
  windows:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1043,11 +1043,6 @@ jobs:
          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
-
      - name: Pack artifacts
        id: pack_artifacts
        run: |
@@ -1083,6 +1078,11 @@ jobs:
          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+
  windows-sycl:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1142,11 +1142,6 @@ jobs:
            -DLLAMA_BUILD_BORINGSSL=ON
          cmake --build build --target ggml-sycl -j %NUMBER_OF_PROCESSORS%

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-x64-sycl
-
      - name: Build the release package
        id: pack_artifacts
        run: |
@@ -1193,6 +1188,11 @@ jobs:
          path: llama-bin-win-sycl-x64.zip
          name: llama-bin-win-sycl-x64.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-x64-sycl
+
  ubuntu-24-sycl:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1265,11 +1265,6 @@ jobs:
            -DGGML_SYCL_F16=${{ matrix.fp16 }}
          time cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -1286,6 +1281,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
          name: llama-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
+
 #   ubuntu-22-rocm:
 #     needs: [check-release, get-version]
 #     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1380,11 +1380,6 @@ jobs:
 #             ${{ env.CMAKE_ARGS }}
 #           cmake --build build --config Release -j $(nproc)

-#       # - name: ccache-clear
-#       #   uses: ./.github/actions/ccache-clear
-#       #   with:
-#       #     key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
-
 #       - name: Determine tag name
 #         id: tag
 #         uses: ./.github/actions/get-tag-name
@@ -1404,6 +1399,11 @@ jobs:
 #           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
 #           name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz

+#       # - name: ccache-clear
+#       #   uses: ./.github/actions/ccache-clear
+#       #   with:
+#       #     key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
+
  ios-xcode:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1688,6 +1688,7 @@ jobs:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        with:
          tag_name: ${{ steps.tag.outputs.name }}
+          prerelease: true
          body: |
            <details open>

@@ -4,8 +4,8 @@ include(CheckIncludeFileCXX)

 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 1)
-set(LLAMA_VERSION_PATCH 2)
+set(LLAMA_VERSION_MINOR 2)
+set(LLAMA_VERSION_PATCH 0)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")

 # whether this is a development/nightly build
@@ -116,7 +116,7 @@ in inline assembler.
 Most kernels are very naive with lots of low hanging fruits left:

 > [!IMPORTANT]
-> Several assembly instructions emmited by the compiler are not implemented
+> Several assembly instructions emitted by the compiler are not implemented
 > in hardware and software emulation in firmware is not ready yet.
 > Eventually firmware will transparently trap unimplemented instructions
 > and will emulate them inside exception handler. Until then, kernel
@@ -138,12 +138,12 @@ Most kernels are very naive with lots of low hanging fruits left:
 > kernel build process. Feel free to take ideas/code from there or try linking
 > it in.

-Before commiting any changes to operations and/or kernels, don't forget
+Before committing any changes to operations and/or kernels, don't forget
 to update supported ops reports (instructions at `docs/ops.md`).

 When logging is enabled (e.g. by setting `--log-file` cli param),
 each compute kernel run outputs a line with
-pipe-delimited key-value pairs containing kernel level performance infomation.
+pipe-delimited key-value pairs containing kernel level performance information.
 Line is prefixed with `ET_PERF`:

 ```
@@ -160,7 +160,7 @@ to `GGML_ET_PROFILE/et_runtime_trace.json` and `GGML_ET_PROFILE/kernel_map` on e

 ### Uberkernel

-The in-knernel implementaiton of device dispatch/kernel fusion. The ET SDK has a non-trivial op-to-op gap. `Uberkernel` (name taken from the original Esperanto AI's compiler)
+The in-kernel implementation of device dispatch/kernel fusion. The ET SDK has a non-trivial op-to-op gap. `Uberkernel` (name taken from the original Esperanto AI's compiler)
 dispatches multiple already existing kernel implementations with device side synchronization. Due to the processor's design, there is no natural memory visibility
 horizon between sub-kernel invocations. This makes uberkernel much more difficult to develop and debug. Currently Uberkerel is hidden begind the
 `GGML_ET_UBERKERNEL` environment variable and is disabled by default. Setting it to 1 enables it and provides significant performance improvements but is only
@@ -4,8 +4,8 @@ project("ggml" C CXX ASM)

 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 20)
-set(GGML_VERSION_PATCH 2)
+set(GGML_VERSION_MINOR 21)
+set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -639,6 +639,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)

        set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
@@ -701,6 +702,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa_asm.S
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla_asm.S
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c
@@ -737,8 +740,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
        set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
    endif()

-    if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
-        # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
-        target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
-    endif()
+  if (CMAKE_C_COMPILER_ID STREQUAL "IntelLLVM" OR CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+      # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
+      target_compile_options(${GGML_CPU_NAME} PRIVATE "$<$<OR:$<COMPILE_LANG_AND_ID:C,IntelLLVM>,$<COMPILE_LANG_AND_ID:CXX,IntelLLVM>>:$<$<BOOL:${WIN32}>:/clang:>-fno-associative-math>")
+  endif()
+
 endfunction()
@@ -23,6 +23,7 @@
 #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.h"
 #include "kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.h"
 #include "kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.h"
+#include "kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.h"
 #include "kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.h"

 #include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
@@ -76,6 +77,21 @@ static inline void kernel_run_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
    Fn(m, n, k, lhs, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
 }

+template <void (*Fn)(size_t, size_t, size_t, const void *, size_t, const void *, void *, size_t, size_t, float, float)>
+static inline void kernel_run_lhs_stride_fn10(size_t       m,
+                                              size_t       n,
+                                              size_t       k,
+                                              size_t       lhs_stride,
+                                              const void * lhs,
+                                              const void * rhs,
+                                              void *       dst,
+                                              size_t       dst_stride_row,
+                                              size_t       dst_stride_col,
+                                              float        clamp_min,
+                                              float        clamp_max) {
+    Fn(m, n, k, lhs, lhs_stride, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
+}
+
 template<void(*Fn)(size_t,size_t,size_t,const void*,const void*,float*,size_t,size_t,float,float)>
 static inline void kernel_run_float_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
                                         const void* lhs, const void* rhs, void* dst,
@@ -947,25 +963,25 @@ static ggml_kleidiai_kernels ggml_kleidiai_kernels_f32[] = {
            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_f32p2vlx1_f32_sme>,
            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_f32p2vlx1_f32_sme>,
        },
-        /* SME GEMV */
+        /* SME2 GEMV */
        {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_lhs_offset_ex     = */ nullptr,
-            /* .get_rhs_packed_offset_ex = */ nullptr,
-            /* .run_kernel_ex         = */ nullptr,
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_mr                = */ kai_get_m_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
+            /* .run_kernel_ex         = */ &kernel_run_lhs_stride_fn10<kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
        },
        /* .gemv_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_f32p2vlx1_f32_sme,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_pack_f32p2vlx1_f32_sme>,
-            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_f32p2vlx1_f32_sme>,
-            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_f32p2vlx1_f32_sme>,
+            /* .get_offset            = */ nullptr,
+            /* .get_packed_offset_ex  = */ nullptr,
+            /* .packed_size_ex        = */ nullptr,
+            /* .pack_func_ex          = */ nullptr,
        },
        /* .rhs_info = */ {
            /* .packed_stride         = */ nullptr,
@@ -696,6 +696,15 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        }

        if (op->src[0]->type == GGML_TYPE_F32) {
+            ggml_kleidiai_kernels * primary     = kernel_chain[0];
+            kernel_info *           gemv_kernel = primary ? &primary->gemv : nullptr;
+            if (is_gemv && op->src[1]->nb[0] == (int64_t) sizeof(float) && gemv_kernel &&
+                gemv_kernel->get_lhs_offset_ex && gemv_kernel->get_rhs_packed_offset_ex &&
+                gemv_kernel->run_kernel_ex && gemv_kernel->get_dst_offset) {
+                size = 0;
+                return true;
+            }
+
            size_t cursor = 0;
            bool any_slot = false;

@@ -811,15 +820,28 @@ class tensor_traits : public ggml::cpu::tensor_traits {
            return false;
        }

-        kernel_info * kernel        = &kernels->gemm;
+        const size_t k = ne00;
+        const size_t m = ne11;
+        const size_t n = ne01;
+        const bool use_gemv = m == 1 && src1->nb[0] == (int64_t) sizeof(float) &&
+                              kernels->gemv.get_lhs_offset_ex &&
+                              kernels->gemv.get_rhs_packed_offset_ex &&
+                              kernels->gemv.run_kernel_ex &&
+                              kernels->gemv.get_dst_offset;
+
+        kernel_info * kernel        = use_gemv ? &kernels->gemv : &kernels->gemm;
        lhs_packing_info * lhs_info = &kernels->gemm_lhs_info;

-        if (!kernel || !lhs_info || !lhs_info->get_offset || !lhs_info->get_packed_offset_ex ||
-            !lhs_info->packed_size_ex || !lhs_info->pack_func_ex ||
+        if (!kernel || !kernel->get_lhs_offset_ex ||
            !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex || !kernel->get_dst_offset) {
            return false;
        }

+        if (!use_gemv && (!lhs_info || !lhs_info->get_offset || !lhs_info->get_packed_offset_ex ||
+                          !lhs_info->packed_size_ex || !lhs_info->pack_func_ex)) {
+            return false;
+        }
+
        const kleidiai_weight_header * header = kleidiai_weight_header_from_ptr(src0->data);
        const bool has_header = kleidiai_is_weight_header_valid(header);

@@ -832,16 +854,14 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        const int nth = params->nth > 0 ? params->nth : 1;
        const int ith = params->ith;

-        const size_t k = ne00;
-        const size_t m = ne11;
-        const size_t n = ne01;
-
        const size_t mr = kernel->get_mr();
        const size_t kr = kernel->get_kr();
        const size_t sr = kernel->get_sr();

-        const size_t lhs_packed_size = lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
-        GGML_ASSERT(lhs_packed_size <= params->wsize);
+        const size_t lhs_packed_size = use_gemv ? 0 : lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
+        if (!use_gemv) {
+            GGML_ASSERT(lhs_packed_size <= params->wsize);
+        }

        uint8_t * lhs_packed   = static_cast<uint8_t *>(params->wdata);
        const size_t dst_stride = dst->nb[1];
@@ -853,7 +873,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
            const uint8_t * lhs_batch_base = static_cast<const uint8_t *>(src1->data) + batch_idx * src1->nb[2];
            uint8_t * dst_batch_base = static_cast<uint8_t *>(dst->data) + batch_idx * dst->nb[2];

-            {
+            if (!use_gemv) {
                const int64_t m_roundup_mr = kai_roundup((int64_t)m, (int64_t)mr);
                int64_t max_threads = mr ? (m_roundup_mr / (int64_t)mr) : nth;
                max_threads = std::max<int64_t>(1, max_threads);
@@ -903,15 +923,17 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                const size_t n_to_process = std::min(chunk_cols, n - n_start);

                if (n_to_process > 0) {
-                    const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
+                    const size_t lhs_offset = use_gemv ? kernel->get_lhs_offset_ex(0, k, 0)
+                                                       : lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
                    const size_t rhs_packed_offset = kernel->get_rhs_packed_offset_ex(n_start, k, 0);
                    const size_t dst_offset        = kernel->get_dst_offset(0, n_start, dst_stride);

-                    const void * lhs_ptr = lhs_packed + lhs_packed_offset;
+                    const void * lhs_ptr = use_gemv ? lhs_batch_base + lhs_offset
+                                                    : lhs_packed + lhs_offset;
                    const void * rhs_ptr = rhs_base + rhs_packed_offset;
                    float * dst_ptr      = reinterpret_cast<float *>(dst_batch_base + dst_offset);

-                    kernel->run_kernel_ex(m, n_to_process, k, 0,
+                    kernel->run_kernel_ex(m, n_to_process, k, use_gemv ? src1->nb[1] : 0,
                                          lhs_ptr,
                                          rhs_ptr,
                                          dst_ptr,
@@ -3180,8 +3180,9 @@ static bool ggml_hexagon_supported_argsort(const struct ggml_hexagon_session * s
 static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
    const int32_t * op_params = &op->op_params[0];

-    if (op_params[15] != 0) {
-        return false; // FIXME: support ggml_rope_set_offset
+    // ggml_rope_set_offset: HVX kernels need a VLEN-aligned window start (32 f32 elems)
+    if (op_params[15] % 32 != 0) {
+        return false;
    }

    int mode = op_params[2];
@@ -53,6 +53,7 @@

 struct htp_rope_context {
    int32_t n_dims;
+    int32_t n_offs;
    int32_t mode;
    int32_t n_ctx_orig;
    int32_t sections[4];
@@ -405,32 +406,40 @@ static inline void hvx_rope_f32_aa(float * restrict dst, const float * restrict

 static void inline rope_basic_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
                   uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    const uint32_t n_offs = rctx->n_offs; // VLEN-aligned (enforced by supports_op)
    #pragma unroll(4)
    for (uint32_t i = 0; i < nr; i++) {
        float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
        float * s = (float *) (src + i * rctx->src0_row_size_aligned);

-        hvx_rope_f32_aa(d, s, rctx->n_dims, theta_cache);
+        hvx_rope_f32_aa(d + n_offs, s + n_offs, rctx->n_dims, theta_cache);

        // fill the remain channels with data from src tensor
-        if (rctx->n_dims < ne0) {
-            hvx_copy_f32_uu((uint8_t *)(d + rctx->n_dims), (uint8_t *)(s + rctx->n_dims), ne0 - rctx->n_dims);
+        if (n_offs > 0) {
+            hvx_copy_f32_uu((uint8_t *) d, (uint8_t *) s, n_offs);
+        }
+        if (n_offs + rctx->n_dims < ne0) {
+            hvx_copy_f32_uu((uint8_t *)(d + n_offs + rctx->n_dims), (uint8_t *)(s + n_offs + rctx->n_dims), ne0 - n_offs - rctx->n_dims);
        }
    }
 }

 static void inline rope_neox_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
                   uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    const uint32_t n_offs = rctx->n_offs; // VLEN-aligned (enforced by supports_op)
    #pragma unroll(4)
    for (uint32_t i = 0; i < nr; i++) {
        float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
        float * s = (float *) (src + i * rctx->src0_row_size_aligned);

-        hvx_rope_neox_f32_aa(d, s, rctx->n_dims, theta_cache);
+        hvx_rope_neox_f32_aa(d + n_offs, s + n_offs, rctx->n_dims, theta_cache);

        // fill the remain channels with data from src tensor
-        if (rctx->n_dims < ne0) {
-            hvx_copy_f32_uu((uint8_t *)(d + rctx->n_dims), (uint8_t *)(s + rctx->n_dims), ne0 - rctx->n_dims);
+        if (n_offs > 0) {
+            hvx_copy_f32_uu((uint8_t *) d, (uint8_t *) s, n_offs);
+        }
+        if (n_offs + rctx->n_dims < ne0) {
+            hvx_copy_f32_uu((uint8_t *)(d + n_offs + rctx->n_dims), (uint8_t *)(s + n_offs + rctx->n_dims), ne0 - n_offs - rctx->n_dims);
        }
    }
 }
@@ -673,6 +682,7 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
    rctx.n_dims     = ((const int32_t *) op_params)[1];
    rctx.mode       = ((const int32_t *) op_params)[2];
    rctx.n_ctx_orig = ((const int32_t *) op_params)[4];
+    rctx.n_offs     = ((const int32_t *) op_params)[15];

    memcpy(&rctx.freq_base,   (int32_t *) op_params + 5,  sizeof(float));
    memcpy(&rctx.freq_scale,  (int32_t *) op_params + 6,  sizeof(float));
@@ -10365,9 +10365,12 @@ kernel void kernel_mul_mm(
    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));

    // Configure matmul operation
+    // note: K is dynamic_extent (clamped to the valid range in PHASE 2), since a static
+    //       N_MM_NK_TOTAL K tile would read src1 out of bounds when K % N_MM_NK_TOTAL != 0
+    // ref: https://github.com/ggml-org/llama.cpp/pull/27064
    mpp::tensor_ops::matmul2d<
        mpp::tensor_ops::matmul2d_descriptor(
-            NRB, NRA, N_MM_NK_TOTAL, false, true, true,
+            NRB, NRA, static_cast<int>(dynamic_extent), false, true, true,
            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;

@@ -10419,10 +10422,14 @@ kernel void kernel_mul_mm(
        threadgroup_barrier(mem_flags::mem_threadgroup);

        // === PHASE 2: Tensor matmul ===
-        auto mA = tA.slice(0, 0);
-        auto mB = tB.slice(loop_k, rb);
+        // Clamp the K extent of both operand tensors to the remaining valid K range so
+        // the dynamic-K op never reads past the K extent of src1 (or the staged A tile).
+        const int kExt = min(N_MM_NK_TOTAL, K - loop_k);

-        mm.run(mB, mA, cT);
+        auto tAv = tensor(sa, dextents<int32_t, 2>(kExt, NRA), array<int, 2>({1, N_MM_NK_TOTAL}));
+        auto tBv = tensor(ptrB + loop_k + rb * strideB, dextents<int32_t, 2>(kExt, N - rb), array<int, 2>({1, strideB}));
+
+        mm.run(tBv, tAv, cT);

        threadgroup_barrier(mem_flags::mem_threadgroup);
    }
@@ -582,6 +582,8 @@ struct ggml_backend_opencl_context {
    bool adreno_use_bin_kernels;
    get_adreno_bin_kernel_func_t get_adreno_bin_kernel_func = nullptr;
    ggml_cl_compiler_version adreno_cl_compiler_version;
+    // The q6_K flat mul_mat codegen workarounds are needed by old E031 compilers only.
+    bool q6_k_flat_old_compiler;

    std::string kernel_compile_opts;  // cached for lazy-compiled kernels.

@@ -1931,8 +1933,14 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
 #else
        const std::string kernel_src = read_file("mul_mv_q6_k_f32_flat.cl");
 #endif
+        // The codegen workarounds in this kernel are a measured 13-20% loss on
+        // compilers that do not need them, so only the affected ones build them;
+        // everyone else gets the original source.
+        const std::string q6k_opts = backend_ctx->q6_k_flat_old_compiler
+            ? compile_opts + " -DADRENO_OLD_COMPILER=1"
+            : compile_opts;
        cl_program prog =
-            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+            build_program_from_source(backend_ctx, kernel_src.c_str(), q6k_opts);

        CL_CHECK((backend_ctx->kernel_mul_mv_q6_K_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q6_K_f32_flat", &err), err));
        CL_CHECK(clReleaseProgram(prog));
@@ -5917,6 +5925,16 @@ static ggml_backend_opencl_context * ggml_cl_init(ggml_backend_dev_t dev) {
        (backend_ctx->adreno_cl_compiler_version.type == E031 && backend_ctx->adreno_cl_compiler_version.major >= 47) ||
        (backend_ctx->adreno_cl_compiler_version.type == DX   && backend_ctx->adreno_cl_compiler_version.major >= 17);

+    // The q6_K flat mul_mat miscompile is a defect of the older E031 compilers, not a
+    // property of any GPU generation: it reproduces on E031.38 (Adreno 642L) and E031.41
+    // (Adreno 740) and is fixed by E031.45 (Adreno 619). Gate on the compiler so parts
+    // that do not need the workarounds do not pay for them. The explicit type check is
+    // required: newer_than_or_same() is false for every non-E031 compiler, so negating it
+    // alone would enable the workarounds on E17/DX.
+    backend_ctx->q6_k_flat_old_compiler =
+        backend_ctx->adreno_cl_compiler_version.type == E031 &&
+        !backend_ctx->adreno_cl_compiler_version.newer_than_or_same(E031, 45, 0, 0);
+
    size_t ext_str_size;
    clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
    char *ext_buffer = (char *)alloca(ext_str_size + 1);
@@ -7375,6 +7393,19 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                       op->src[0]->type == GGML_TYPE_Q4_K  ||
                       op->src[0]->type == GGML_TYPE_Q5_K  ||
                       op->src[0]->type == GGML_TYPE_Q6_K) {
+                // The E031.41 compiler (usually with A7x) miscompiles the flat K-quant
+                // GEMV kernels (kernel_mul_mv_q*_K_f32_flat) and makes lm_head run much
+                // slower than it should. So, make it fallback to CPU to preserve performance
+                // for this compiler series.
+                static const char * a7x_lmhead_env = getenv("GGML_OPENCL_A7X_LMHEAD_CPU");
+                static const bool   a7x_lmhead_cpu = (a7x_lmhead_env == nullptr || a7x_lmhead_env[0] != '0');
+                if (a7x_lmhead_cpu &&
+                    backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X &&
+                    (op->src[0]->type == GGML_TYPE_Q4_K || op->src[0]->type == GGML_TYPE_Q5_K ||
+                     op->src[0]->type == GGML_TYPE_Q6_K) &&
+                    op->src[0]->ne[1] >= 32768) {   // vocab-scale weight; no FFN/attn weight is this tall
+                    return false;
+                }
                return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
            } else if (op->src[0]->type == GGML_TYPE_Q8_0) {
                return op->src[1]->type == GGML_TYPE_F32;
@@ -7416,9 +7447,6 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
        case GGML_OP_DIAG_MASK_INF:
            return op->ne[3] == 1;
        case GGML_OP_ROPE: {
-            if (((const int32_t *) op->op_params)[15] != 0) {
-                return false; // FIXME: support ggml_rope_set_offset
-            }
            const int mode = ((const int32_t *) op->op_params)[2];
            const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
            const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
@@ -7496,6 +7524,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                                     v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16;
            const bool is_f32_f16  = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 &&
                                     v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F32;
+
            const bool is_f32_q8_0 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_Q8_0 &&
                                     v->type == GGML_TYPE_Q8_0 && op->type == GGML_TYPE_F32 &&
                                     dk % 32 == 0 && dv % 32 == 0;
@@ -7503,6 +7532,21 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                                     v->type == GGML_TYPE_Q4_0 && op->type == GGML_TYPE_F32 &&
                                     dk % 32 == 0 && dv % 32 == 0;

+            // A7X (Adreno 740, compiler E031.41) SIGSEGVs inside clBuildProgram
+            // building the flash_attn programs whose KV path is mixed-type or
+            // dequantized — f32_f16, q8_0, q4_0 (reproduced at DK=40 and DK=64; it
+            // is DK-independent). It is a driver crash, not codegen-wrong-output, so
+            // it cannot be caught in-process (fatal=false only handles clean compile
+            // errors). The uniform f16_f16 / f32_f32 programs compile fine on this
+            // compiler, so decline only the KV-convert variants; ggml then runs
+            // those (f16-KV / quant-KV) attention layers on the CPU backend.
+            // Negative compiler carve-out, same idiom as the Intel DK=512 decline
+            // below and the X1E driver-quirk guards.
+            if (backend_ctx && backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X &&
+                (is_f32_f16 || is_f32_q8_0 || is_f32_q4_0)) {
+                return false;
+            }
+
            // Asymmetric KV: host-dequants both sides to F32, uses f32 kernel.
            auto is_kv_type_ok = [](ggml_type t) {
                return t == GGML_TYPE_F16 || t == GGML_TYPE_F32 ||
@@ -12830,7 +12874,10 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const
    GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
    GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);

-    const int nth = MIN(64, ne00);
+    int nth = 1;
+    while (nth < ne00 && nth < 64) {
+        nth *= 2;
+    }

    cl_kernel kernel = backend_ctx->kernel_norm;

@@ -20580,6 +20627,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne1));
            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &r2));
            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r3));
+            // The optimizer-barrier arg exists only in the ADRENO_OLD_COMPILER build of
+            // this kernel; conformant compilers get the original 17-arg signature.
+            if (backend_ctx->q6_k_flat_old_compiler) {
+                cl_uchar q6k_mask = 0xFF;   // never 0xFE in prod; see the kernel note
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_uchar), &q6k_mask));
+            }
 #else
            kernel = backend_ctx->kernel_mul_mv_q6_K_f32;

@@ -23867,6 +23920,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
    const int n_dims     = ((int *) dst->op_params)[1];
    const int mode       = ((int *) dst->op_params)[2];
    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    const int n_offs     = ((int32_t *) dst->op_params)[15];

    float freq_base;
    float freq_scale;
@@ -23895,6 +23949,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const

    if (is_vision) {
        GGML_ASSERT(n_dims == ne00/2);
+        GGML_ASSERT(n_offs == 0); // offset not supported for vision, as the rotated pairs span the whole row
    }

    cl_kernel kernel;
@@ -23986,6 +24041,12 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
    if (is_mrope && !is_vision) {
        CL_CHECK(clSetKernelArg(kernel, 34, sizeof(int), &is_imrope));
    }
+    // norm and neox have n_offs after beta_slow, mrope has it after is_imrope
+    if (!is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int), &n_offs));
+    } else if (is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 35, sizeof(int), &n_offs));
+    }

    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
    size_t local_work_size[] = {(size_t)nth, 1, 1};
@@ -28,6 +28,13 @@

 #define QK_K       256

+// ADRENO_OLD_COMPILER is defined by the host (-D) only for the Adreno E031
+// compilers older than E031.45, which miscompile several constructs this kernel
+// used (confirmed on E031.38 and E031.41; E031.45 is clean). Every other
+// compiler -- newer E031, E17, DX, Intel, and every non-Adreno device that
+// builds this program -- takes the #else branches, which are the original
+// source: the workarounds below cost ~13% on the q6_K flat n=1 GEMV where they
+// are not needed.
 inline float block_q_6_K_dot_y_flat(
    global uchar * blk_ql,
    global uchar * blk_qh,
@@ -37,6 +44,9 @@ inline float block_q_6_K_dot_y_flat(
    int ip,
    int is,
    int l0,
+#if defined(ADRENO_OLD_COMPILER)
+    int dbg,
+#endif
    float4 y0,
    float4 y1,
    float4 y2,
@@ -48,10 +58,40 @@ inline float block_q_6_K_dot_y_flat(
    global uchar * q1 = blk_ql     + ib*128 + q_offset_l;
    global uchar * q2 = q1         + QK_K/8;
    global uchar * qh = blk_qh     + ib*64 + q_offset_h;
-    global char  * sc = blk_scales + ib*16 + is;

    float dall = blk_d[ib];

+#if defined(ADRENO_OLD_COMPILER)
+    // The vectorized dequant (int4/float4 bit-ops, convert_*4, dot()) and vload4
+    // are miscompiled here -> garbage weights. Reconstruct the 6-bit weights and
+    // take the dot product scalar. q4_K/q5_K flat already use scalar paths, which
+    // is why q6_K was the only flat GEMV that failed.
+    // Scales are SIGNED int8; read as uchar and sign-extend arithmetically so the
+    // result does not depend on whether the compiler treats `char` as signed.
+    global uchar * sc = (global uchar *)(blk_scales + ib*16 + is);
+
+    int s0 = (int)sc[0] - 256*(sc[0] >> 7);
+    int s2 = (int)sc[2] - 256*(sc[2] >> 7);
+    int s4 = (int)sc[4] - 256*(sc[4] >> 7);
+    int s6 = (int)sc[6] - 256*(sc[6] >> 7);
+
+    // one 6-bit weight: low/high nibble of a ql byte OR'd with a 2-bit qh plane
+    // (plane p in {0,1,2,3} selects qh bits 2p..2p+1) placed at bits 4-5, minus 32.
+    #define Q6W(qb, sh, hb, p) ((float)((((int)(qb) >> (sh)) & 15) | ((((int)(hb) >> (2*(p))) & 3) << 4)) - 32.f)
+
+    float d0 = y0.s0*Q6W(q1[0],0,qh[0],0) + y0.s1*Q6W(q1[1],0,qh[1],0) + y0.s2*Q6W(q1[2],0,qh[2],0) + y0.s3*Q6W(q1[3],0,qh[3],0);
+    float d1 = y1.s0*Q6W(q2[0],0,qh[0],1) + y1.s1*Q6W(q2[1],0,qh[1],1) + y1.s2*Q6W(q2[2],0,qh[2],1) + y1.s3*Q6W(q2[3],0,qh[3],1);
+    float d2 = y2.s0*Q6W(q1[0],4,qh[0],2) + y2.s1*Q6W(q1[1],4,qh[1],2) + y2.s2*Q6W(q1[2],4,qh[2],2) + y2.s3*Q6W(q1[3],4,qh[3],2);
+    float d3 = y3.s0*Q6W(q2[0],4,qh[0],3) + y3.s1*Q6W(q2[1],4,qh[1],3) + y3.s2*Q6W(q2[2],4,qh[2],3) + y3.s3*Q6W(q2[3],4,qh[3],3);
+    #undef Q6W
+
+    if (dbg) printf("HELPER dall=%f s=[%d %d %d %d] d=[%f %f %f %f] ql0=%d qh0=%d y00=%f\n",
+        dall, s0, s2, s4, s6, d0, d1, d2, d3, (int)q1[0], (int)qh[0], y0.s0);
+
+    return dall * (d0 * s0 + d1 * s2 + d2 * s4 + d3 * s6);
+#else
+    global char * sc = blk_scales + ib*16 + is;
+
    // Vectorized loads: 3 uchar4 weight loads instead of 12 scalar byte reads.
    // q_offset_l/h are 4-aligned, so these are aligned vector loads.
    uchar4 q1v = vload4(0, q1);
@@ -72,6 +112,7 @@ inline float block_q_6_K_dot_y_flat(

    return dall * (dot(y0, w0) * sc[0] + dot(y1, w1) * sc[2] +
                   dot(y2, w2) * sc[4] + dot(y3, w3) * sc[6]);
+#endif
 }

 #undef N_DST
@@ -113,6 +154,11 @@ kernel void kernel_mul_mv_q6_K_f32_flat(
        int ne1,
        int r2,
        int r3
+#if defined(ADRENO_OLD_COMPILER)
+        ,
+        uchar q6k_mask   // runtime 0xFF; the host passes it so the compiler cannot
+                         // constant-fold the printf guards below into nothing
+#endif
 ) {
    src1 = (global float*)((global char*)src1 + offset1);
    dst = (global float*)((global char*)dst + offsetd);
@@ -128,6 +174,22 @@ kernel void kernel_mul_mv_q6_K_f32_flat(

    int first_row = (N_SIMDGROUP * r0 + get_sub_group_id()) * N_DST;

+#if defined(ADRENO_OLD_COMPILER)
+    // 64-bit `ulong` integer arithmetic is miscompiled here -> the base-pointer byte
+    // offsets came out wrong, so EVERY weight/scale read hit the wrong address. This
+    // was the primary cause of the q6_K flat failure (q5_K uses int offsets and is
+    // unaffected). Compute the block index in `int` and widen to `ulong` only inside
+    // the pointer expression: the byte offset stays 64-bit, but there is no ulong
+    // arithmetic chain to miscompile. The int index would overflow past ~2^31 blocks,
+    // which no realistic weight reaches -- but that is a narrowing, so keep it off the
+    // conformant path, which retains full ulong arithmetic.
+    int offset_src0 = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global uchar * blk_ql     = (global uchar *) src0_ql + (ulong)offset_src0 * 128;
+    global uchar * blk_qh     = (global uchar *) src0_qh + (ulong)offset_src0 * 64;
+    global char  * blk_scales = (global char  *) src0_s  + (ulong)offset_src0 * 16;
+    global half  * blk_d      = (global half  *) src0_d  + offset_src0;
+#else
    ulong offset_src0    = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
    ulong offset_src0_ql = offset_src0 * 128;
    ulong offset_src0_qh = offset_src0 * 64;
@@ -138,6 +200,7 @@ kernel void kernel_mul_mv_q6_K_f32_flat(
    global uchar * blk_qh     = (global uchar *) src0_qh + offset_src0_qh;
    global char  * blk_scales = (global char  *) src0_s  + offset_src0_s;
    global half  * blk_d      = (global half  *) src0_d  + offset_src0_d;
+#endif
    global float * yy         = (global float *) src1    + r1*ne10 + im*ne00*ne1;

    int tid = get_sub_group_local_id()%(N_SIMDWIDTH/BLOCK_STRIDE); // within-super-block part, 0..15
@@ -155,24 +218,55 @@ kernel void kernel_mul_mv_q6_K_f32_flat(

    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
        global float * y = yy + ib * QK_K + 128*ip + l0;
+#if defined(ADRENO_OLD_COMPILER)
+        // vload4 of f32 is miscompiled here; index the lanes scalar instead.
+        float4 y0 = (float4)(y[ 0], y[ 1], y[ 2], y[ 3]);
+        float4 y1 = (float4)(y[32], y[33], y[34], y[35]);
+        float4 y2 = (float4)(y[64], y[65], y[66], y[67]);
+        float4 y3 = (float4)(y[96], y[97], y[98], y[99]);
+#else
        float4 y0 = vload4(0, y +  0);
        float4 y1 = vload4(0, y + 32);
        float4 y2 = vload4(0, y + 64);
        float4 y3 = vload4(0, y + 96);
+#endif

        for (int row = 0; row < N_DST; row++) {
            if (first_row + row < ne01) {
+#if defined(ADRENO_OLD_COMPILER)
+                int dbg = (q6k_mask==0xFE && r0==0 && r1==0 && im==0 && row==0 && ib==0 &&
+                           ne00==256 && ne01==16 && get_sub_group_local_id()==0) ? 1 : 0;
+                sumf[row] += block_q_6_K_dot_y_flat(
+                    blk_ql + row*nb*128, blk_qh + row*nb*64, blk_scales + row*nb*16, blk_d + row*nb,
+                    ib, ip, is, l0, dbg, y0, y1, y2, y3);
+#else
                sumf[row] += block_q_6_K_dot_y_flat(
                    blk_ql + row*nb*128, blk_qh + row*nb*64, blk_scales + row*nb*16, blk_d + row*nb,
                    ib, ip, is, l0, y0, y1, y2, y3);
+#endif
            }
        }
    }

+#if defined(ADRENO_OLD_COMPILER)
+    // Optimizer barrier. This compiler drops the sumf partials unless a side effect
+    // forces them to materialize. q6k_mask is a kernel arg the compiler cannot prove
+    // is never 0xFE (the host always passes 0xFF), so the printf survives compilation
+    // but never executes. FRAGILE: the exact set and placement of these guarded
+    // printfs is load-bearing on E031.41 -- removing any one re-breaks q6_K.
+    if (q6k_mask==0xFE && r0==0 && r1==0 && im==0 && ne00==256 && ne01==16 && get_sub_group_local_id()<16) {
+        printf("Q6KLANE lane=%d ip=%d il=%d is=%d l0=%d sumf0=%f\n",
+            get_sub_group_local_id(), ip, il, is, l0, sumf[0]);
+    }
+#endif
    for (int row = 0; row < N_DST; row++) {
        float tot = sub_group_reduce_add(sumf[row]);
        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+#if defined(ADRENO_OLD_COMPILER)
+            if (q6k_mask==0xFE && r0==0 && r1==0 && im==0 && row==0 && ne00==256 && ne01==16)
+                printf("Q6KTOT tot=%f\n", tot);
+#endif
        }
    }
 }
@@ -75,7 +75,8 @@ kernel void kernel_rope_norm_f32(
        float ext_factor,
        float attn_factor,
        float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -94,14 +95,15 @@ kernel void kernel_rope_norm_f32(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

-            float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            float theta = theta_base * pow(freq_base, inv_ndims*iw);

            float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

            global float * src       = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -154,7 +156,8 @@ kernel void kernel_rope_norm_f16(
        float ext_factor,
        float attn_factor,
        float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -173,14 +176,15 @@ kernel void kernel_rope_norm_f16(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

-            float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            float theta = theta_base * pow(freq_base, inv_ndims*iw);

            float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -233,7 +237,8 @@ kernel void kernel_rope_neox_f32(
        float ext_factor,
        float attn_factor,
        float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -252,17 +257,18 @@ kernel void kernel_rope_neox_f32(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(freq_base, inv_ndims*iw);

            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

-            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);

            const float x0 = src[0];
            const float x1 = src[n_dims/2];
@@ -312,7 +318,8 @@ kernel void kernel_rope_neox_f16(
        float ext_factor,
        float attn_factor,
        float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -331,17 +338,18 @@ kernel void kernel_rope_neox_f16(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(freq_base, inv_ndims*iw);

            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

-            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);

            const float x0 = src[0];
            const float x1 = src[n_dims/2];
@@ -393,7 +401,8 @@ kernel void kernel_rope_multi_f32(
        float beta_fast,
        float beta_slow,
        int4 sections,
-        int  is_imrope
+        int  is_imrope,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -414,10 +423,11 @@ kernel void kernel_rope_multi_f32(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

-            const int sector = (i0 / 2) % sect_dims;
+            const int sector = ic % sect_dims;
            float theta_base = 0.0f;

            if (is_imrope) {
@@ -445,14 +455,14 @@ kernel void kernel_rope_multi_f32(
                }
            }

-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(freq_base, inv_ndims*iw);

            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

-            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);

            const float x0 = src[0];
            const float x1 = src[n_dims/2];
@@ -504,7 +514,8 @@ kernel void kernel_rope_multi_f16(
        float beta_fast,
        float beta_slow,
        int4 sections,
-        int  is_imrope
+        int  is_imrope,
+        int n_offs
 ) {
    src0 = (global void*)((global char*)src0 + offset0);
    src1 = (global int*)((global char*)src1 + offset1);
@@ -525,10 +536,11 @@ kernel void kernel_rope_multi_f16(
    float inv_ndims = -1.f/n_dims;

    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;

-            const int sector = (i0 / 2) % sect_dims;
+            const int sector = ic % sect_dims;
            float theta_base = 0.0f;

            if (is_imrope) {
@@ -556,14 +568,14 @@ kernel void kernel_rope_multi_f16(
                }
            }

-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(freq_base, inv_ndims*iw);

            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;

-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);

-            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);

            const float x0 = src[0];
            const float x1 = src[n_dims/2];
@@ -1955,6 +1955,23 @@ static void dequantize_mul_mat_vec_q4_K_sycl_reorder_esimd(const void *vx, const
    });
 }

+static void dequantize_mul_mat_vec_q5_K_sycl_reorder_esimd(const void *vx, const float *y,
+                                                           float *dst, const int ncols,
+                                                           const int nrows,
+                                                           dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int workgroups = (nrows + 1) / 2;
+    stream->submit([&](sycl::handler &h) {
+        sycl::local_accessor<float, 1> lmem(sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE * 2), h);
+        h.parallel_for(
+            sycl::nd_range<1>(sycl::range<1>((size_t)workgroups * GGML_SYCL_DMMV_ESIMD_WG_SIZE), sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE)),
+            [=](sycl::nd_item<1> it) [[intel::sycl_explicit_simd]] {
+                dequantize_mul_mat_vec_reorder_esimd<GGML_TYPE_Q5_K>(
+                    vx, y, dst, ncols, nrows, lmem, it);
+            });
+    });
+}
+
 static void dequantize_mul_mat_vec_q6_K_sycl_reorder_esimd(const void *vx, const float *y,
                                                           float *dst, const int ncols,
                                                           const int nrows,
@@ -2134,7 +2151,15 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
        case GGML_TYPE_Q5_K:
            if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
                ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
-                dequantize_mul_mat_vec_q5_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+#ifdef GGML_SYCL_DMMV_HAS_ESIMD
+                if (g_ggml_sycl_enable_esimd) {
+                    dequantize_mul_mat_vec_q5_K_sycl_reorder_esimd(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
+                else
+#endif
+                {
+                    dequantize_mul_mat_vec_q5_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
            } else {
                dequantize_mul_mat_vec_q5_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
            }
@@ -62,7 +62,7 @@

 #define DPCT_UNUSED(x) (void)(x)

-inline void _abort(const char * str) {
+[[noreturn]] inline void _abort(const char * str) {
    std::cerr << str << std::endl;
    std::abort();
 }
@@ -10,7 +10,7 @@
    (ITEM.get_local_range(IDX) * ITEM.get_group(IDX) + ITEM.get_local_id(IDX))

 static void acc_f32(const char * x, const char * y, float * dst, const int64_t ne,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2,
    const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
    const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
@@ -455,7 +455,7 @@ static void unary_mul_sycl(const T * x, const T * g, T * dst, const int64_t k, c
 namespace ggml_sycl_detail {
 static void acc_f32_sycl(const char *x, const char *y, float *dst,
                         const int64_t n_elements,
-                         const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+                         const int64_t ne0, const int64_t ne1, const int64_t ne2,
                         const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
                         const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
                         const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
@@ -466,7 +466,7 @@ static void acc_f32_sycl(const char *x, const char *y, float *dst,
                                           sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
                         [=](sycl::nd_item<3> /*item_ct1*/) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                            acc_f32(x, y, dst, n_elements,
-                                ne0, ne1, ne2, ne3,
+                                ne0, ne1, ne2,
                                nb00, nb01, nb02, nb03,
                                ne10, ne11, ne12, ne13,
                                nb10, nb11, nb12, nb13,
@@ -970,7 +970,7 @@ static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor
    const int64_t offset = (int64_t) ((const int32_t *) dst->op_params)[3] / (int64_t) sizeof(float);

    ggml_sycl_detail::acc_f32_sycl(src0_d, src1_d, dst_d, ggml_nelements(dst),
-        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2],
        src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
        src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
        src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
@@ -1,15 +1,3 @@
-//
-// MIT license
-// Copyright (C) 2026 Intel Corporation
-// SPDX-License-Identifier: MIT
-//
-
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-
 #ifndef GGML_SYCL_ESIMD_HPP
 #define GGML_SYCL_ESIMD_HPP

@@ -287,6 +275,128 @@ template <> struct esimd_reorder_q_traits<GGML_TYPE_Q4_K> {
    }
 };

+// ---------------------------------------------------------------------------
+// Q5_K, SOA reorder layout produced by reorder_qw_q5_k:
+//   [qs: nb*(QK_K/2)] [qh: nb*(QK_K/8)] [scales: nb*K_SCALE_SIZE] [dm: nb*sizeof(half2)]
+// with nb = nrows*num_blocks_per_row.
+//
+// Identical to Q4_K except each 4-bit quant gains a 5th (high) bit from qh:
+// output chunk c (0..7) adds 16 when bit c of qh[l] is set, where qh[l] indexes
+// the same 32 bytes for every chunk (matches dequantize_row_q5_K).
+// ---------------------------------------------------------------------------
+template <> struct esimd_reorder_q_traits<GGML_TYPE_Q5_K> {
+    struct ptrs {
+        const uint8_t *    qs;
+        const uint8_t *    qh;
+        const uint8_t *    scales;
+        const sycl::half * dm;
+    };
+
+    static ESIMD_INLINE ptrs make_ptrs(const void * vx, size_t nb) {
+        const uint8_t * qs     = (const uint8_t *) vx;
+        const uint8_t * qh     = qs + nb * (QK_K / 2);
+        const uint8_t * scales = qh + nb * (QK_K / 8);
+        const sycl::half * dm  = (const sycl::half *) (scales + nb * K_SCALE_SIZE);
+        return { qs, qh, scales, dm };
+    }
+
+    // extract bit `bit` (0..7) of each lane and move it to bit position 4,
+    // e.g. for the 4-bit base quant's 5th (high) bit. `bit` is always a
+    // compile-time-known unrolled loop constant at call sites, so this folds
+    // to a single mask (bit==4), mask+left-shift (bit<4), or mask+right-shift
+    // (bit>4) instead of the shift+mask+shift a naive `(qh>>bit & 1) << 4` emits.
+    static ESIMD_INLINE sycl::ext::intel::esimd::simd<uint16_t, 32> extract_bit_to_pos4(
+            sycl::ext::intel::esimd::simd<uint8_t, 32> qh, int bit) {
+        using namespace sycl::ext::intel::esimd;
+        simd<uint16_t, 32> masked = convert<uint16_t>(qh & simd<uint8_t, 32>((uint8_t) (1u << bit)));
+        if (bit < 4) {
+            return masked << simd<uint16_t, 32>((uint16_t) (4 - bit));
+        } else if (bit > 4) {
+            return masked >> simd<uint16_t, 32>((uint16_t) (bit - 4));
+        }
+        return masked;
+    }
+
+    static ESIMD_INLINE void mac_pair(
+            const ptrs & pa, size_t bia,
+            const ptrs & pb, size_t bib, bool has_b,
+            sycl::ext::intel::esimd::simd<float, 256> & y_vec,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_a,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_b) {
+        using namespace sycl::ext::intel::esimd;
+
+        simd<uint8_t, 128> qs_a     = block_load<uint8_t, 128>(pa.qs + bia * (QK_K / 2));
+        simd<uint8_t, 128> qs_b     = 0;
+        simd<uint8_t, 32>  qh_a     = block_load<uint8_t, 32>(pa.qh + bia * (QK_K / 8));
+        simd<uint8_t, 32>  qh_b     = 0;
+        simd<uint8_t, 12>  scales_a = block_load<uint8_t, 12>(pa.scales + bia * K_SCALE_SIZE);
+        simd<uint8_t, 12>  scales_b = 0;
+
+        const float dall_a = (float) pa.dm[bia * 2 + 0];
+        const float dmin_a = (float) pa.dm[bia * 2 + 1];
+        float dall_b = 0.0f;
+        float dmin_b = 0.0f;
+        if (has_b) {
+            qs_b     = block_load<uint8_t, 128>(pb.qs + bib * (QK_K / 2));
+            qh_b     = block_load<uint8_t, 32>(pb.qh + bib * (QK_K / 8));
+            scales_b = block_load<uint8_t, 12>(pb.scales + bib * K_SCALE_SIZE);
+            dall_b = (float) pb.dm[bib * 2 + 0];
+            dmin_b = (float) pb.dm[bib * 2 + 1];
+        }
+
+        simd<float, 8> scale_f_a, min_f_a, scale_f_b, min_f_b;
+        unpack_scale_min_k4(scales_a, dall_a, dmin_a, scale_f_a, min_f_a);
+        unpack_scale_min_k4(scales_b, dall_b, dmin_b, scale_f_b, min_f_b);
+
+        simd<uint8_t, 128> qs_lo_a = qs_a & simd<uint8_t, 128>(0x0F);
+        simd<uint8_t, 128> qs_hi_a = qs_a >> simd<uint8_t, 128>(4);
+        simd<uint8_t, 128> qs_lo_b = qs_b & simd<uint8_t, 128>(0x0F);
+        simd<uint8_t, 128> qs_hi_b = qs_b >> simd<uint8_t, 128>(4);
+
+#pragma unroll
+        for (int sb = 0; sb < 8; sb += 2) {
+            const int q_offset = sb * 16;
+            simd<float, 32> y_lo = y_vec.select<32, 1>(sb * 32);
+            simd<float, 32> y_hi = y_vec.select<32, 1>((sb + 1) * 32);
+
+            const float scale_a_lo = scale_f_a[sb];
+            const float scale_a_hi = scale_f_a[sb + 1];
+            const float min_a_lo   = min_f_a[sb];
+            const float min_a_hi   = min_f_a[sb + 1];
+            const float scale_b_lo = scale_f_b[sb];
+            const float scale_b_hi = scale_f_b[sb + 1];
+            const float min_b_lo   = min_f_b[sb];
+            const float min_b_hi   = min_f_b[sb + 1];
+
+            simd<uint8_t, 32> qa_lo_u8 = qs_lo_a.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qa_hi_u8 = qs_hi_a.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qb_lo_u8 = qs_lo_b.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qb_hi_u8 = qs_hi_b.select<32, 1>(q_offset);
+            simd<uint16_t, 32> qa_lo = convert<uint16_t>(qa_lo_u8);
+            simd<uint16_t, 32> qa_hi = convert<uint16_t>(qa_hi_u8);
+            simd<uint16_t, 32> qb_lo = convert<uint16_t>(qb_lo_u8);
+            simd<uint16_t, 32> qb_hi = convert<uint16_t>(qb_hi_u8);
+
+            // add the 5th bit: chunk sb uses qh bit sb, chunk sb+1 uses qh bit sb+1;
+            // qh always indexes the same 32 bytes regardless of chunk
+            qa_lo += extract_bit_to_pos4(qh_a, sb);
+            qa_hi += extract_bit_to_pos4(qh_a, sb + 1);
+            qb_lo += extract_bit_to_pos4(qh_b, sb);
+            qb_hi += extract_bit_to_pos4(qh_b, sb + 1);
+
+            simd<float, 32> deq_a_lo = convert<float>(qa_lo) * scale_a_lo + min_a_lo;
+            simd<float, 32> deq_a_hi = convert<float>(qa_hi) * scale_a_hi + min_a_hi;
+            simd<float, 32> deq_b_lo = convert<float>(qb_lo) * scale_b_lo + min_b_lo;
+            simd<float, 32> deq_b_hi = convert<float>(qb_hi) * scale_b_hi + min_b_hi;
+
+            acc_a += y_lo * deq_a_lo;
+            acc_b += y_lo * deq_b_lo;
+            acc_a += y_hi * deq_a_hi;
+            acc_b += y_hi * deq_b_hi;
+        }
+    }
+};
+
 // ---------------------------------------------------------------------------
 // Q6_K, SOA reorder layout:
 //   [ql: nb*(QK_K/2)] [qh: nb*(QK_K/4)] [scales(int8): nb*(QK_K/16)] [d: nb*half]
@@ -43,7 +43,7 @@ static void mkl_fa_pack_q_fp16(
    dpct::queue_ptr stream,
    sycl::half * __restrict dst,
    const float * __restrict q_src,
-    int n_queries, int n_query_rows, int DKQ,
+    int n_queries, int DKQ,
    int gqa_ratio, int kvh_base_head,
    float q_scale, int64_t q_row_stride, int64_t q_head_stride,
    int64_t wg_size) {
@@ -121,7 +121,7 @@ static void mkl_fa_online_softmax_chunk(
    float * __restrict VKQ_accum,
    int q0, int q_rows, int n_queries, int DV,
    int chunk_size, int chunk_start,
-    int kvh_head, int gqa_ratio,
+    int kvh_head,
    const sycl::half * mask_data, int64_t mask_head_stride,
    int64_t mask_row_stride, int mask_n_heads,
    float logit_softcap, int64_t wg_size) {
@@ -473,7 +473,6 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *
    MKL_ACCUM(dequant_time_us, t_deq);

    // --- Resolve mask pointers ---
-    const sycl::half * mask_data = nullptr;
    int64_t mask_head_stride = 0;
    int64_t mask_row_stride  = 0;
    int     mask_n_heads     = 0;
@@ -547,7 +546,7 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *
            // 1. Pack all GQA Q heads into fp16 (full n_query_rows)
            mkl_fa_pack_q_fp16(stream,
                Q_head_f16_ptr, Q_batch,
-                n_queries, n_query_rows, DKQ,
+                n_queries, DKQ,
                gqa_ratio, kvh_base_head,
                q_scale, q_row_stride, q_head_stride, wg_size);

@@ -605,7 +604,7 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *
                            KQ_max_ptr, KQ_sum_ptr, VKQ_accum_ptr,
                            q0, q_rows, n_queries, DV,
                            this_chunk, chunk_start,
-                            kvh_base_head, gqa_ratio,
+                            kvh_base_head,
                            mask_batch, mask_head_stride,
                            mask_row_stride, mask_n_heads,
                            logit_softcap, wg_size);
@@ -21,14 +21,6 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
    if (!g_ggml_sycl_fa_onednn) {
        return false;
    }
-    // Battlemage (Xe2) only, for now. On other Intel archs oneDNN's fused SDPA returns wrong results
-    // for some shapes (e.g. head_dim=64 on Arc / xe_hpg) -- an oneDNN bug tracked upstream at
-    // https://github.com/uxlfoundation/oneDNN/issues/5510. Remove this hardware limitation once that
-    // is fixed; until then non-BMG archs fall back to the existing FA kernel.
-    const gpu_arch arch = ggml_sycl_info().devices[ggml_sycl_get_device()].hw_info.arch;
-    if (arch != gpu_arch::intel_gpu_bmg_g21 && arch != gpu_arch::intel_gpu_bmg_g31) {
-        return false;
-    }
    const ggml_tensor * Q     = dst->src[0];
    const ggml_tensor * K     = dst->src[1];
    const ggml_tensor * V     = dst->src[2];
@@ -60,6 +52,17 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
            }
        }
    }
+    // This is the improved SPDA gate. Rather than gating Alchemist GPUs from all SPDA features, we instead target only the failing shapes.
+    // If the GPU being assessed isn't in the grouping below, it has full access to all SPDA shapes. Otherwise, if it's an Alchemist GPU, we block only the shapes with head sizes that fail.
+    // It is much easier to compare the device to a small list of failing cases than to define all the passing ones.
+    const gpu_arch arch = ggml_sycl_info().devices[ggml_sycl_get_device()].hw_info.arch;
+    bool support_spda = !(arch == gpu_arch::intel_gpu_dg2_g10 ||
+                         arch == gpu_arch::intel_gpu_dg2_g11 ||
+                         arch == gpu_arch::intel_gpu_dg2_g12);
+
+    if (!support_spda && K->ne[0] == 64) {
+        return false;
+    }
    // Optional KV-length ceiling (GGML_SYCL_FA_ONEDNN_MAX_KV, 0 = unlimited). Escape hatch:
    // very long sequences make the fused SDPA slow enough to risk the xe driver watchdog on
    // some stacks; past the cap we fall back to the native FA kernel instead.
@@ -921,16 +921,16 @@ ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,

    void * dev_ptr;
    if (use_usm_system) {
-        GGML_SYCL_DEBUG("[SYCL] allocating %lu Bytes with USM system\n", size);
+        GGML_SYCL_DEBUG("[SYCL] allocating %zu Bytes with USM system\n", size);
        dev_ptr = (void *)aligned_malloc_host(alignment, aligned_size);
        if (!dev_ptr) {
-            GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on host\n", __func__, size);
+            GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on host\n", __func__, size);
            return nullptr;
        }
    } else {
        SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)ggml_sycl_malloc_device(size, *stream)));
        if (!dev_ptr) {
-          GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
+          GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on device\n", __func__, size);
          return nullptr;
        }
    }
@@ -1177,7 +1177,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
        SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)ggml_sycl_malloc_device(size, *stream)));
        if (!buf) {
            char err_buf[1024];
-            snprintf(err_buf, 1023, "%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
+            snprintf(err_buf, 1023, "%s: can't allocate %zu Bytes of memory on device\n", __func__, size);
            throw std::runtime_error(err_buf);
        }
        // set padding to 0 to avoid possible NaN values
@@ -1517,8 +1517,13 @@ static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggm
 }

 static size_t ggml_backend_sycl_host_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
-    ggml_backend_sycl_device_context * dev_ctx = (ggml_backend_sycl_device_context *) buft->device->context;
-    return dpct::dev_mgr::instance().get_device(dev_ctx->device).get_max_mem_alloc_size();
+
+    if (g_ggml_sycl_enable_host_pinned_mem) {
+        ggml_backend_sycl_device_context * dev_ctx = (ggml_backend_sycl_device_context *) buft->device->context;
+        return dpct::dev_mgr::instance().get_device(dev_ctx->device).get_max_mem_alloc_size();
+    } else {
+        return SIZE_MAX;
+    }
 }

 ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
@@ -1646,7 +1651,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {

        SYCL_CHECK(CHECK_TRY_ERROR(ptr = (void *)ggml_sycl_malloc_device(look_ahead_size, *qptr)));
        if (!ptr) {
-            GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device/GPU\n", __func__, look_ahead_size);
+            GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on device/GPU\n", __func__, look_ahead_size);
            return nullptr;
        }

@@ -1658,7 +1663,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
                (uint32_t)(max_size/1024/1024), (uint32_t)(g_sycl_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024));
 #endif

-        // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%lu, return %p\n", look_ahead_size, ptr);
+        // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%zu, return %p\n", look_ahead_size, ptr);
        return ptr;
    }

@@ -1838,7 +1843,7 @@ struct ggml_sycl_pool_host : public ggml_sycl_pool {

            SYCL_CHECK(CHECK_TRY_ERROR(ptr = (void *) sycl::malloc_host(size, *qptr)));
            if (!ptr) {
-                GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on host\n", __func__, size);
+                GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on host\n", __func__, size);
                return nullptr;
            }
            pool_size += size;
@@ -2774,9 +2779,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
        const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();

        {
+#if GGML_SYCL_DNNL
            const int64_t gemm_flops = (int64_t)row_diff * src1_ncols * ne10;
            const bool use_mkl_direct = gemm_flops < 256 * 256 * 256;
-#if GGML_SYCL_DNNL
            if (g_ggml_sycl_enable_dnn && !use_mkl_direct) {
                DnnlGemmWrapper::row_gemm(ctx, row_diff, src1_ncols, ne10, src0_ddf_i,
                                          DnnlGemmWrapper::to_dt<float>(), src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
@@ -3513,7 +3518,9 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
    float *            dst_ddf  = static_cast<float *>(dst->data);

    const sycl::half * src1_f16       = static_cast<const sycl::half *>(src1->data);
+#if GGML_SYCL_DNNL
    const size_t       type_size_src0 = ggml_type_size(src0->type);
+#endif
    const size_t       type_size_src1 = ggml_type_size(src1->type);

    bool is_src0_cont_2 = ggml_is_contiguous_2(src0);
@@ -3530,6 +3537,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
        scope_op_debug_print    scope_dbg_print(__func__, "/to_fp16_nc_sycl", dst, /*num_src=*/2,
                                                " : converting src1 to fp16");

+#if GGML_SYCL_DNNL
        // iterate tensor dims and find the slowest moving dim and stride
        int last_dim=0;
        int last_str=0;
@@ -3549,7 +3557,6 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
            }

        }
-#if GGML_SYCL_DNNL
        // oneDNN handles strided data and does not need overhead of ggml_get_to_fp16_nc_sycl
        const int64_t ne_src1 = src1->nb[last_str] * src1->ne[last_dim] / type_size_src1;
        src1_f16_alloc.alloc(ne_src1);
@@ -3804,6 +3811,7 @@ static bool ggml_sycl_supports_reorder_esimd(enum ggml_type type) {
    switch (type) {
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
        case GGML_TYPE_Q6_K:
            return true;
        default:
@@ -6242,8 +6250,6 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
        }
        case GGML_OP_ROPE:
        case GGML_OP_ROPE_BACK:
-            // FIXME: support ggml_rope_set_offset
-            return ((const int32_t *) op->op_params)[15] == 0;
        case GGML_OP_IM2COL:
        case GGML_OP_IM2COL_3D:
        case GGML_OP_UPSCALE:
@@ -85,7 +85,7 @@ static void im2col_sycl(const float *   x,
    */
    stream->parallel_for(sycl::nd_range<3>(block_nums * sycl::range<3>(1, 1, MIN(IC_KH_KW, SYCL_IM2COL_BLOCK_SIZE)),
                                           sycl::range<3>(1, 1, MIN(IC_KH_KW, SYCL_IM2COL_BLOCK_SIZE))),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3>) {
                             im2col_kernel(x, dst, IC, IW, IH, OH, OW, KW, KH, IC_IH_IW, IH_IW, N_OH, KH_KW, IC_KH_KW,
                                           s0, s1, p0, p1, d0, d1);
                         });
@@ -271,7 +271,7 @@ static void im2col_3d_sycl(const float *   src,
    */
    stream->parallel_for(sycl::nd_range<3>(block_nums * sycl::range<3>(1, 1, MIN(IC_KD_KH_KW, SYCL_IM2COL_BLOCK_SIZE)),
                                           sycl::range<3>(1, 1, MIN(IC_KD_KH_KW, SYCL_IM2COL_BLOCK_SIZE))),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3>) {
                             im2col_3d_kernel(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW, OH_OW, KD_KH_KW,
                                              ID_IH_IW, KH_KW, IH_IW, IC_ID_IH_IW, IC_KD_KH_KW, OW_KD_KH_KW,
                                              OD_OH_OW_IC_KD_KH_KW, OH_OW_IC_KD_KH_KW, OW_IC_KD_KH_KW, N_OD_OH, OD_OH,
@@ -7,9 +7,6 @@ static void norm_f32(const float* x, float* dst, const int ncols,
    const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
    const float eps, const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {

-    const int nrows = item_ct1.get_group_range(2);
-    const int nchannels = item_ct1.get_group_range(1);
-
    const int nthreads = item_ct1.get_local_range(2);
    const int sample  = item_ct1.get_group(0);
    const int channel = item_ct1.get_group(1);
@@ -155,9 +152,6 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols,
    const float* mul = nullptr, const int64_t mul_stride_row = 0, const int64_t mul_stride_channel = 0,
    const int64_t mul_stride_sample = 0, const int mul_nrows = 0, const int mul_nchannels = 0, const int mul_nsamples = 0) {

-    const int nrows = item_ct1.get_group_range(2);
-    const int nchannels = item_ct1.get_group_range(1);
-
    const int sample  = item_ct1.get_group(0);
    const int channel = item_ct1.get_group(1);
    const int row     = item_ct1.get_group(2);
@@ -225,8 +219,6 @@ static void l2_norm_f32(const float * x, float * dst, const int ncols,
    const int64_t src_stride_sample, const int64_t dst_stride_col, const int64_t dst_stride_row,
    const int64_t dst_stride_channel, const int64_t dst_stride_sample, const float eps,
    const sycl::nd_item<3>& item_ct1, float* s_sum, const int block_size) {
-    const int nrows     = item_ct1.get_group_range(2);
-    const int nchannels = item_ct1.get_group_range(1);

    const int row     = item_ct1.get_group(2);
    const int channel = item_ct1.get_group(1);
@@ -41,7 +41,7 @@ template <bool forward, bool has_ff, typename T, typename D>
 static void rope_norm(const T *x, D *dst, const int ne00, const int ne01,
                      const int ne02, const int s01, const int s02,
                      const int s03, const int s1, const int s2, const int s3,
-                      const int n_dims, const int32_t *pos,
+                      const int n_dims, const int n_offs, const int32_t *pos,
                      const float freq_scale, const float ext_factor,
                      const float attn_factor, const rope_corr_dims corr_dims,
                      const float theta_scale, const float *freq_factors,
@@ -78,19 +78,21 @@ static void rope_norm(const T *x, D *dst, const int ne00, const int ne01,
            ggml_sycl_memcpy_1<4>(dst + idst, &v);
        }
    };
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
        store_coaelsced(x[ix + 0], x[ix + 1]);
        return;
    }

-    const float theta_base = pos[i2] * dpct::pow(theta_scale, i0 / 2.0f);
+    const int iw = i0 - n_offs; // relative idx

-    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
+    const float theta_base = pos[i2] * dpct::pow(theta_scale, iw / 2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[iw / 2] : 1.0f;

    float cos_theta;
    float sin_theta;

-    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, i0,
+    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, iw,
                       ext_factor, attn_factor, cos_theta, sin_theta);

    const float x0 = x[ix + 0];
@@ -104,7 +106,7 @@ template <bool forward, bool has_ff, typename T, typename D>
 static void rope_neox(const T *x, D *dst, const int ne00, const int ne01,
                      const int ne02, const int s01, const int s02,
                      const int s03, const int s1, const int s2, const int s3,
-                      const int n_dims, const int32_t *pos,
+                      const int n_dims, const int n_offs, const int32_t *pos,
                      const float freq_scale, const float ext_factor,
                      const float attn_factor, const rope_corr_dims corr_dims,
                      const float theta_scale, const float *freq_factors,
@@ -132,35 +134,38 @@ static void rope_neox(const T *x, D *dst, const int ne00, const int ne01,
        idst += row_indices[i2] * set_rows_stride;
    }

-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
        dst[idst + i0 / 2 + 0] = ggml_sycl_cast<D>(x[ix + i0 / 2 + 0]);
        dst[idst + i0 / 2 + 1] = ggml_sycl_cast<D>(x[ix + i0 / 2 + 1]);

        return;
    }

-    const float theta_base = pos[i2] * dpct::pow(theta_scale, i0 / 2.0f);
+    const int iw = i0 - n_offs; // relative idx

-    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
+    const float theta_base = pos[i2] * dpct::pow(theta_scale, iw / 2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[iw / 2] : 1.0f;

    float cos_theta;
    float sin_theta;

-    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, i0,
+    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, iw,
                       ext_factor, attn_factor, cos_theta, sin_theta);

-    const float x0 = x[ix + 0];
-    const float x1 = x[ix + n_dims / 2];
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is n_offs + iw/2 = i0/2 + n_offs/2
+    const float x0 = x[ix + n_offs / 2 + 0];
+    const float x1 = x[ix + n_offs / 2 + n_dims / 2];

-    dst[idst + 0] = ggml_sycl_cast<D>(x0 * cos_theta - x1 * sin_theta);
-    dst[idst + n_dims / 2] = ggml_sycl_cast<D>(x0 * sin_theta + x1 * cos_theta);
+    dst[idst + n_offs / 2 + 0] = ggml_sycl_cast<D>(x0 * cos_theta - x1 * sin_theta);
+    dst[idst + n_offs / 2 + n_dims / 2] = ggml_sycl_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }

 template <bool forward, bool has_ff, typename T>
 static void rope_multi(const T *x, T *dst, const int ne00, const int ne01,
                       const int ne02, const int s01, const int s02,
                       const int s03, const int s1, const int s2, const int s3,
-                       const int n_dims, const int32_t *pos,
+                       const int n_dims, const int n_offs, const int32_t *pos,
                       const float freq_scale, const float ext_factor,
                       const float attn_factor, const rope_corr_dims corr_dims,
                       const float theta_scale, const float *freq_factors,
@@ -183,54 +188,57 @@ static void rope_multi(const T *x, T *dst, const int ne00, const int ne01,
    int idst = i0 / 2 + i1 * s1 + i2 * s2 + i3 * s3;
    const int ix = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;

-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
        dst[idst + i0 / 2 + 0] = x[ix + i0 / 2 + 0];
        dst[idst + i0 / 2 + 1] = x[ix + i0 / 2 + 1];

        return;
    }

+    const int iw = i0 - n_offs; // relative idx
+
    const int sect_dims =
        sections.v[0] + sections.v[1] + sections.v[2] + sections.v[3];
    const int sec_w = sections.v[1] + sections.v[0];
-    const int sector = (i0 / 2) % sect_dims;
+    const int sector = (iw / 2) % sect_dims;

    float theta_base = 0.0;
    if (is_imrope) {
        if (sector % 3 == 1 && sector < 3 * sections.v[1]) { // h
-            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, iw / 2.0f);
        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) { // w
-            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, iw / 2.0f);
        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) { // t
-            theta_base = pos[i2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * dpct::pow(theta_scale, iw / 2.0f);
        } else {
-            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, iw / 2.0f);
        }
    } else {
        if (sector < sections.v[0]) {
-            theta_base = pos[i2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * dpct::pow(theta_scale, iw / 2.0f);
        } else if (sector >= sections.v[0] && sector < sec_w) {
-            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, iw / 2.0f);
        } else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, iw / 2.0f);
        } else if (sector >= sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, iw / 2.0f);
        }
    }

-    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
+    const float freq_factor = has_ff ? freq_factors[iw / 2] : 1.0f;

    float cos_theta;
    float sin_theta;

-    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, i0,
+    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, iw,
                       ext_factor, attn_factor, cos_theta, sin_theta);

-    const float x0 = x[ix + 0];
-    const float x1 = x[ix + n_dims / 2];
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is n_offs + iw/2 = i0/2 + n_offs/2
+    const float x0 = x[ix + n_offs / 2 + 0];
+    const float x1 = x[ix + n_offs / 2 + n_dims / 2];

-    dst[idst + 0] = x0 * cos_theta - x1 * sin_theta;
-    dst[idst + n_dims / 2] = x0 * sin_theta + x1 * cos_theta;
+    dst[idst + n_offs / 2 + 0] = x0 * cos_theta - x1 * sin_theta;
+    dst[idst + n_offs / 2 + n_dims / 2] = x0 * sin_theta + x1 * cos_theta;
 }

 template <bool forward, bool has_ff, typename T>
@@ -293,7 +301,7 @@ static void
 rope_norm_sycl(const T *x, D *dst, const int ne00, const int ne01,
               const int ne02, const int s01, const int s02, const int s03,
               const int s1, const int s2, const int s3, const int n_dims,
-               const int nr, const int32_t *pos, const float freq_scale,
+               const int n_offs, const int nr, const int32_t *pos, const float freq_scale,
               const float freq_base, const float ext_factor,
               const float attn_factor, const rope_corr_dims corr_dims,
               const float *freq_factors, const int64_t *row_indices,
@@ -313,7 +321,7 @@ rope_norm_sycl(const T *x, D *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_norm<forward, false>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, row_indices, set_rows_stride);
            });
    } else {
@@ -323,7 +331,7 @@ rope_norm_sycl(const T *x, D *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_norm<forward, true>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, row_indices, set_rows_stride);
            });
    }
@@ -334,7 +342,7 @@ static void
 rope_neox_sycl(const T *x, D *dst, const int ne00, const int ne01,
               const int ne02, const int s01, const int s02, const int s03,
               const int s1, const int s2, const int s3, const int n_dims,
-               const int nr, const int32_t *pos, const float freq_scale,
+               const int n_offs, const int nr, const int32_t *pos, const float freq_scale,
               const float freq_base, const float ext_factor,
               const float attn_factor, const rope_corr_dims corr_dims,
               const float *freq_factors, const int64_t *row_indices,
@@ -354,7 +362,7 @@ rope_neox_sycl(const T *x, D *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_neox<forward, false>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, row_indices, set_rows_stride);
            });
    } else {
@@ -364,7 +372,7 @@ rope_neox_sycl(const T *x, D *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_neox<forward, true>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, row_indices, set_rows_stride);
            });
    }
@@ -375,7 +383,7 @@ static void
 rope_multi_sycl(const T *x, T *dst, const int ne00, const int ne01,
                const int ne02, const int s01, const int s02, const int s03,
                const int s1, const int s2, const int s3, const int n_dims,
-                const int nr, const int32_t *pos, const float freq_scale,
+                const int n_offs, const int nr, const int32_t *pos, const float freq_scale,
                const float freq_base, const float ext_factor,
                const float attn_factor, const rope_corr_dims corr_dims,
                const float *freq_factors, const mrope_sections sections,
@@ -395,7 +403,7 @@ rope_multi_sycl(const T *x, T *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_multi<forward, false, T>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, sections, is_imrope);
            });
    } else {
@@ -405,7 +413,7 @@ rope_multi_sycl(const T *x, T *dst, const int ne00, const int ne01,
                GGML_UNUSED(item_ct1);
                rope_multi<forward, true, T>(
                    x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                    theta_scale, freq_factors, sections, is_imrope);
            });
    }
@@ -497,6 +505,7 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
    const int n_dims = ((int32_t *)dst->op_params)[1];
    const int mode = ((int32_t *)dst->op_params)[2];
    const int n_ctx_orig = ((int32_t *)dst->op_params)[4];
+    const int n_offs = ((int32_t *)dst->op_params)[15];
    mrope_sections sections;

    float freq_base;
@@ -526,6 +535,7 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,

    if (is_vision) {
        GGML_ASSERT(n_dims == ne00 / 2);
+        GGML_ASSERT(n_offs == 0); // offset not supported for vision, as the rotated pairs span the whole row
    }

    const int32_t *pos = (const int32_t *)src1_d;
@@ -545,19 +555,19 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
            rope_neox_sycl<forward, float, float>(
                (const float *)src0_d, (float *)dst_d, ne00, ne01, ne02, s01,
-                s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
                set_rows_stride, stream);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
            rope_neox_sycl<forward, float, sycl::half>(
                (const float *)src0_d, (sycl::half *)dst_d, ne00, ne01, ne02,
-                s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                row_indices, set_rows_stride, stream);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
            rope_neox_sycl<forward, sycl::half, sycl::half>(
                (const sycl::half *)src0_d, (sycl::half *)dst_d, ne00, ne01,
-                ne02, s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                row_indices, set_rows_stride, stream);
        } else {
@@ -568,13 +578,13 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
        if (src0->type == GGML_TYPE_F32) {
            rope_multi_sycl<forward>((const float *)src0_d, (float *)dst_d,
                                     ne00, ne01, ne02, s01, s02, s03, s1, s2,
-                                     s3, n_dims, nr, pos, freq_scale, freq_base,
+                                     s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                     ext_factor, attn_factor, corr_dims,
                                     freq_factors, sections, is_imrope, stream);
        } else if (src0->type == GGML_TYPE_F16) {
            rope_multi_sycl<forward>(
                (const sycl::half *)src0_d, (sycl::half *)dst_d, ne00, ne01,
-                ne02, s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                sections, is_imrope, stream);
        } else {
@@ -602,19 +612,19 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
            rope_norm_sycl<forward, float, float>(
                (const float *)src0_d, (float *)dst_d, ne00, ne01, ne02, s01,
-                s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
                set_rows_stride, stream);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
            rope_norm_sycl<forward, float, sycl::half>(
                (const float *)src0_d, (sycl::half *)dst_d, ne00, ne01, ne02,
-                s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                row_indices, set_rows_stride, stream);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
            rope_norm_sycl<forward, sycl::half, sycl::half>(
                (const sycl::half *)src0_d, (sycl::half *)dst_d, ne00, ne01,
-                ne02, s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                row_indices, set_rows_stride, stream);
        } else {
@@ -291,7 +291,7 @@ static void set_rows_sycl(

    stream->parallel_for(
        sycl::nd_range<1>(grid_size * block_size, block_size),
-        [=](sycl::nd_item<1> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+        [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
            k_set_rows<TIn, TIdx, TOut>(
                src0_d, src1_d, dst_d,
                ne00, ne01, ne02,
@@ -2714,6 +2714,7 @@ static webgpu_encoded_op ggml_webgpu_rope(webgpu_context & ctx,
    const int n_dims     = ((int32_t *) dst->op_params)[1];
    const int mode       = ((int32_t *) dst->op_params)[2];
    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    const int n_offs     = ((int32_t *) dst->op_params)[15];

    float freq_base;
    float freq_scale;
@@ -2762,7 +2763,8 @@ static webgpu_encoded_op ggml_webgpu_rope(webgpu_context & ctx,
        (uint32_t) sections[0],
        (uint32_t) sections[1],
        (uint32_t) sections[2],
-        (uint32_t) sections[3]
+        (uint32_t) sections[3],
+        (uint32_t) n_offs
    };

    std::vector<wgpu::BindGroupEntry> entries     = { ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0),
@@ -4472,9 +4474,7 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
            supports_op = (op->type == GGML_TYPE_F32 && src0->type == GGML_TYPE_F32) && ggml_is_contiguous_rows(src0);
            break;
        case GGML_OP_ROPE:
-            // FIXME: support ggml_rope_set_offset
-            supports_op =
-                (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && ((const int32_t *) op->op_params)[15] == 0;
+            supports_op = op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16;
            break;
        case GGML_OP_GLU:
            switch (ggml_get_glu_op(op)) {
@@ -38,7 +38,8 @@ struct Params {
    sections0: u32,
    sections1: u32,
    sections2: u32,
-    sections3: u32
+    sections3: u32,
+    n_offs: u32
 };

@group(0) @binding(0)
@@ -126,7 +127,8 @@ fn rope_yarn(theta_extrap: f32, i: u32) -> vec2<f32> {

 fn pair_base(i0: u32, div_2: bool) -> u32 {
    if (div_2) {
-        return i0 / 2;
+        // first channel of the rotated pair: n_offs + (i0 - n_offs)/2
+        return i0 / 2 + params.n_offs / 2;
    } else {
        return i0;
    }
@@ -165,20 +167,22 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let i_src_row = params.offset_src0 + i3 * params.stride_src03 + i2 * params.stride_src02 + i1 * params.stride_src01;
    let i_dst_row = params.offset_dst + i3 * params.stride_dst3 + i2 * params.stride_dst2 + i1 * params.stride_dst1;

-    if (i0 >= params.n_dims && !is_vision) {
+    if ((i0 < params.n_offs || i0 >= params.n_offs + params.n_dims) && !is_vision) {
        let i_src = i_src_row + i0;
        let i_dst = i_dst_row + i0;
        rotate(i_dst, i_dst + 1, f32(src0[i_src]), f32(src0[i_src + 1]));
        return;
    }

+    let iw = i0 - params.n_offs; // relative idx
+
    var theta_base_mult: u32 = 0;
-    var theta_scale_pwr: u32 = i0 / 2;
+    var theta_scale_pwr: u32 = iw / 2;
    if (is_mrope) {
        let sect_dims = params.sections0 + params.sections1 + params.sections2 + params.sections3;
        let sec_w = params.sections1 + params.sections0;
        let sec_e = params.sections2 + sec_w;
-        let sector = (i0 / 2) % sect_dims;
+        let sector = (iw / 2) % sect_dims;
        if (is_imrope) {
          if (sector % 3 == 1 && sector < 3 * params.sections1) {
              theta_base_mult = 1;
@@ -203,7 +207,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
          } else if (sector >= sec_e) {
              if (is_vision) {
                  theta_scale_pwr = sector - sec_e;
-                  theta_scale_pwr = (i0 / 2) % sec_e;
+                  theta_scale_pwr = (iw / 2) % sec_e;
              }
              theta_base_mult = 3;
          } else if (is_vision) {
@@ -212,7 +216,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
        }
    }
    let theta_base = f32(src1[params.offset_src1 + i2 + params.ne2 * theta_base_mult]) * pow(params.theta_scale, f32(theta_scale_pwr));
-    let thetas = rope_yarn(theta_base/freq_factor(i0), i0);
+    let thetas = rope_yarn(theta_base/freq_factor(iw), iw);

    let i_src = i_src_row + pair_base(i0, is_neox || is_mrope || is_vision);
    let i_dst = i_dst_row + pair_base(i0, is_neox || is_mrope || is_vision);
@@ -15,7 +15,8 @@
 # tag exists.
 #
 # Env (when running in GitHub Actions):
-#   GITHUB_OUTPUT: previous_tag, changelog_title, changelog and nightly are written here
+#   GITHUB_OUTPUT: previous_tag, changelog_title, changelog, nightly and nightly_tag
+#     are written here
 #   GITHUB_REPOSITORY: owner/repo, used to build the nightly release URL (skipped when unset)
 set -euo pipefail

@@ -52,10 +53,10 @@ PREV="$( { git tag --list; echo "${VERSION}"; } \

 if [[ -n "${PREV}" ]]; then
    CHANGELOG="$(git log --oneline "${PREV}..${RELEASE_COMMIT}")"
-    CHANGELOG_TITLE="Change log since ${PREV}"
+    CHANGELOG_TITLE="Changelog since ${PREV}"
 else
    CHANGELOG="(no previous release tag found)"
-    CHANGELOG_TITLE="Change log"
+    CHANGELOG_TITLE="Changelog"
 fi

 # Nightly release: the b* tag pointing at the release commit (|| true: no match is not an error)
@@ -80,6 +81,7 @@ if [[ -n "${GITHUB_OUTPUT:-}" ]]; then
        echo "previous_tag=${PREV}"
        echo "changelog_title=${CHANGELOG_TITLE}"
        echo "nightly=${NIGHTLY}"
+        echo "nightly_tag=${NIGHTLY_TAG}"
        echo "changelog<<CHANGELOG_EOF"
        echo "${CHANGELOG}"
        echo "CHANGELOG_EOF"
@@ -0,0 +1,204 @@
+#!/bin/bash
+#
+# Release preparation script for llama.cpp.
+#
+# Bumps the version in CMakeLists.txt on a release candidate branch.
+# The branch should then be pushed and a PR created, reviewed, and
+# merged. After the PR is merged and the build-cpu workflow has
+# completed successfully, the release is finalized by the make-release
+# workflow (.github/workflows/make-release.yml), which creates the tag.
+#
+# Usage:
+#   ./scripts/release.sh [major|minor|patch] [--dry-run]
+#
+# Example:
+#   $ ./scripts/release.sh minor
+#
+# The script:
+# 1. Creates a release candidate branch (llama-rc-v<major>.<minor>.<patch>)
+# 2. Bumps the version in CMakeLists.txt
+# 3. Commits the version bump
+#
+
+set -e
+
+if [ ! -f "CMakeLists.txt" ] || [ ! -d "scripts" ]; then
+    echo "Error: Must be run from llama.cpp root directory"
+    exit 1
+fi
+
+# Parse command line arguments
+VERSION_TYPE=""
+DRY_RUN=false
+
+for arg in "$@"; do
+    case $arg in
+        --dry-run)
+            DRY_RUN=true
+            ;;
+        major|minor|patch)
+            VERSION_TYPE="$arg"
+            ;;
+        *)
+            echo "Error: Unknown argument '$arg'"
+            echo "Usage: $0 [major|minor|patch] [--dry-run]"
+            exit 1
+            ;;
+    esac
+done
+
+# Default to patch if no version type specified
+VERSION_TYPE="${VERSION_TYPE:-patch}"
+
+# Common validation functions
+check_git_status() {
+    # Check for uncommitted changes (skip in dry-run)
+    if [ "$DRY_RUN" = false ] && ! git diff-index --quiet HEAD --; then
+        echo "Error: You have uncommitted changes. Please commit or stash them first."
+        exit 1
+    fi
+}
+
+check_master_branch() {
+    # Ensure we're on master branch
+    CURRENT_BRANCH=$(git branch --show-current)
+    if [ "$CURRENT_BRANCH" != "master" ]; then
+        if [ "$DRY_RUN" = true ]; then
+            echo "[dry run] Warning: Not on master branch (currently on: $CURRENT_BRANCH). Continuing with dry-run..."
+            echo ""
+        else
+            echo "Error: Must be on master branch. Currently on: $CURRENT_BRANCH"
+            exit 1
+        fi
+    fi
+}
+
+check_master_up_to_date() {
+    # Check if we have the latest from master (skip in dry-run)
+    if [ "$DRY_RUN" = false ]; then
+        echo "Checking if local master is up-to-date with remote..."
+        git fetch origin master
+        LOCAL=$(git rev-parse HEAD)
+        REMOTE=$(git rev-parse origin/master)
+
+        if [ "$LOCAL" != "$REMOTE" ]; then
+            echo "Error: Your local master branch is not up-to-date with origin/master."
+            echo "Please run 'git pull origin master' first."
+            exit 1
+        fi
+        echo "✓ Local master is up-to-date with remote"
+        echo ""
+    elif [ "$(git branch --show-current)" = "master" ]; then
+        echo "[dry run] Warning: Dry-run mode - not checking if master is up-to-date with remote"
+        echo ""
+    fi
+}
+
+# In-place sed that works on both GNU (Linux) and BSD (macOS) sed
+sed_inplace() {
+    if sed --version >/dev/null 2>&1; then
+        sed -i "$@"
+    else
+        sed -i '' "$@"
+    fi
+}
+
+prepare_release() {
+    if [ "$DRY_RUN" = true ]; then
+        echo "[dry-run] Preparing release (no changes will be made)"
+    else
+        echo "Starting release preparation..."
+    fi
+    echo ""
+
+    check_git_status
+    check_master_branch
+    check_master_up_to_date
+
+    # Extract current version from CMakeLists.txt
+    echo "Step 1: Reading current version..."
+    MAJOR=$(grep "set(LLAMA_VERSION_MAJOR" CMakeLists.txt | sed 's/.*MAJOR \([0-9]*\).*/\1/')
+    MINOR=$(grep "set(LLAMA_VERSION_MINOR" CMakeLists.txt | sed 's/.*MINOR \([0-9]*\).*/\1/')
+    PATCH=$(grep "set(LLAMA_VERSION_PATCH" CMakeLists.txt | sed 's/.*PATCH \([0-9]*\).*/\1/')
+
+    echo "Current version: $MAJOR.$MINOR.$PATCH"
+
+    # Calculate new version
+    case $VERSION_TYPE in
+        major)
+            NEW_MAJOR=$((MAJOR + 1))
+            NEW_MINOR=0
+            NEW_PATCH=0
+            ;;
+        minor)
+            NEW_MAJOR=$MAJOR
+            NEW_MINOR=$((MINOR + 1))
+            NEW_PATCH=0
+            ;;
+        patch)
+            NEW_MAJOR=$MAJOR
+            NEW_MINOR=$MINOR
+            NEW_PATCH=$((PATCH + 1))
+            ;;
+    esac
+
+    NEW_VERSION="$NEW_MAJOR.$NEW_MINOR.$NEW_PATCH"
+    RC_BRANCH="llama-rc-v$NEW_VERSION"
+    echo "New release version: $NEW_VERSION"
+    echo "Release candidate branch: $RC_BRANCH"
+    echo ""
+
+    # Create release candidate branch
+    echo "Step 2: Creating release candidate branch..."
+    if [ "$DRY_RUN" = true ]; then
+        echo "  [dry-run] Would create branch: $RC_BRANCH"
+    else
+        git checkout -b "$RC_BRANCH"
+        echo "✓ Created and switched to branch: $RC_BRANCH"
+    fi
+    echo ""
+
+    # Update CMakeLists.txt for release
+    echo "Step 3: Updating version in CMakeLists.txt..."
+    if [ "$DRY_RUN" = true ]; then
+        echo "  [dry-run] Would update LLAMA_VERSION_MAJOR to $NEW_MAJOR"
+        echo "  [dry-run] Would update LLAMA_VERSION_MINOR to $NEW_MINOR"
+        echo "  [dry-run] Would update LLAMA_VERSION_PATCH to $NEW_PATCH"
+    else
+        sed_inplace -e "s/set(LLAMA_VERSION_MAJOR [0-9]*)/set(LLAMA_VERSION_MAJOR $NEW_MAJOR)/" CMakeLists.txt
+        sed_inplace -e "s/set(LLAMA_VERSION_MINOR [0-9]*)/set(LLAMA_VERSION_MINOR $NEW_MINOR)/" CMakeLists.txt
+        sed_inplace -e "s/set(LLAMA_VERSION_PATCH [0-9]*)/set(LLAMA_VERSION_PATCH $NEW_PATCH)/" CMakeLists.txt
+    fi
+    echo ""
+
+    # Commit version bump
+    echo "Step 4: Committing version bump..."
+    if [ "$DRY_RUN" = true ]; then
+        echo "  [dry-run] Would commit: 'llama.cpp : bump version to $NEW_VERSION'"
+    else
+        git add CMakeLists.txt
+        git commit -m "llama.cpp : bump version to $NEW_VERSION"
+    fi
+    echo ""
+
+    echo ""
+    if [ "$DRY_RUN" = true ]; then
+        echo "[dry-run] Summary (no changes were made):"
+        echo "  • Would have created branch: $RC_BRANCH"
+        echo "  • Would have updated version to: $NEW_VERSION"
+    else
+        echo "Release preparation completed!"
+        echo "Summary:"
+        echo "  • Created branch: $RC_BRANCH"
+        echo "  • Updated version to: $NEW_VERSION"
+        echo ""
+        echo "Next steps:"
+        echo "  • Push branch to remote: git push origin $RC_BRANCH"
+        echo "  • Create a Pull Request from $RC_BRANCH to master"
+        echo "  • After the PR is merged and the build-cpu workflow has passed,"
+        echo "    create the release with the make-release workflow"
+        echo "    (.github/workflows/make-release.yml)"
+    fi
+}
+
+prepare_release
@@ -1 +1 @@
-8c63e70982c95ceb862e3a1073a2c1beef75d60a
+8599e0ea3756c4bac4ef813af2241cb1a8bbfb0b
@@ -1062,8 +1062,6 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
        case LLM_ARCH_NEMOTRON_H:
        case LLM_ARCH_NEMOTRON_H_MOE:
        case LLM_ARCH_GRANITE_HYBRID:
-        case LLM_ARCH_LFM2:
-        case LLM_ARCH_LFM2MOE:
        case LLM_ARCH_MINIMAX_01:
        case LLM_ARCH_MINIMAX_M2:
        case LLM_ARCH_MINIMAX_M3:
@@ -487,6 +487,10 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "ssm_out.weight");
        }
        if (std::regex_match(tensor_name, pattern_r_cache) || std::regex_match(tensor_name, pattern_s_cache)) {
+            if (ud->model->arch == LLM_ARCH_LFM2 || ud->model->arch == LLM_ARCH_LFM2MOE) {
+                // the LFM2 shortconv block runs fully mirrored, so its conv state must be mirrored too
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_MIRRORED, "");
+            }
            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0, "ssm_out.weight");
        }
        if (std::regex_match(tensor_name, pattern_ssm_conv1d)) {
@@ -7085,9 +7085,10 @@ struct test_flash_attn_ext : public test_case {
    const ggml_type type_V;
    std::array<int32_t, 4> permute;
    const bool kv_view; // create K/V as views of a larger buffer (like a KV cache)
+    const bool v_is_view_of_k;

    std::string vars() override {
-        return VARS_TO_STR15(hsk, hsv, nh, nr23, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_K, type_V, permute, kv_view);
+        return VARS_TO_STR16(hsk, hsv, nh, nr23, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_K, type_V, permute, kv_view, v_is_view_of_k);
    }

    double max_nmse_err() override {
@@ -7104,9 +7105,9 @@ struct test_flash_attn_ext : public test_case {
    test_flash_attn_ext(int64_t hsk = 128, int64_t hsv = 128, int64_t nh = 32, std::array<int64_t, 2> nr23 = {1, 1}, int64_t kv = 96, int64_t nb = 8,
                        bool mask = true, bool sinks = false, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_prec prec = GGML_PREC_F32,
                        ggml_type type_K = GGML_TYPE_F16, ggml_type type_V = GGML_TYPE_F16, std::array<int32_t, 4> permute = {0, 1, 2, 3},
-                        bool kv_view = true)
+                        bool kv_view = true, bool v_is_view_of_k = false)
        : hsk(hsk), hsv(hsv), nh(nh), nr23(nr23), kv(kv), nb(nb), mask(mask), sinks(sinks), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec),
-          type_K(type_K), type_V(type_V), permute(permute), kv_view(kv_view) {}
+          type_K(type_K), type_V(type_V), permute(permute), kv_view(kv_view), v_is_view_of_k(v_is_view_of_k) {}

    ggml_tensor * build_graph(ggml_context * ctx) override {
        const int64_t hsk_padded = GGML_PAD(hsk, ggml_blck_size(type_K));
@@ -7138,14 +7139,14 @@ struct test_flash_attn_ext : public test_case {
        ggml_set_name(k, "k");

        ggml_tensor * v = nullptr;
-        if (type_K == type_V && hsk_padded == 576 && hsv_padded == 512) {
-            // TODO: this branch should become a separate test case parameter instead of hardcoding this for these head shapes
-
-            // in this branch, the V cache is sub-view of the K cache. this is used by some MLA-based models
+        if (v_is_view_of_k) {
+            // the V cache is a sub-view of the K cache. this is used by some MLA-based models
            // for more info:
            //   - https://github.com/ggml-org/llama.cpp/pull/13435
            //   - https://github.com/ggml-org/llama.cpp/pull/18953#issuecomment-3774948392
            //   - https://github.com/ggml-org/llama.cpp/pull/18986
+            GGML_ASSERT(type_K == type_V && hsv_padded <= hsk_padded);
+
            v = ggml_view_4d(ctx, k, hsv_padded, kv, nh, nr23[1], k->nb[1], k->nb[2], k->nb[3], 0);
        } else {
            v = create_permuted(type_V,        hsv_padded, kv, nh,         nr23[1], kv_view); // the V tensor is usually a view of the V cache
@@ -9298,6 +9299,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {

    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 6, 4096, 5120, {1, 1}, {1, 1}));

+    // K not a multiple of 32
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 64, 32,  65, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 64, 32,  80, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 64, 32,  80, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 32,  80, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 64, 32, 588, {1, 1}, {1, 1})); // 14*14*3, e.g. conv_2d im2col
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 64, 32,  80, {4, 1}, {1, 1}));
+
 #if 0
    // test the mat-mat path for Metal
    for (int k = 1; k < 512; ++k) {
@@ -9898,12 +9907,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                                                    if (hsk != 128 && prec == GGML_PREC_DEFAULT) continue;
                                                    for (ggml_type type_KV : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0, GGML_TYPE_IQ4_NL}) {
                                                        if (type_KV != GGML_TYPE_F16 && hsk != 64 && hsk != 72) continue;
+                                                        // DeepSeek MLA: the V cache is a sub-view of the K cache
+                                                        const bool v_is_view_of_k = hsk == 576;
                                                        test_cases.emplace_back(new test_flash_attn_ext(
-                                                                    hsk, hsv, nh, {nr2, nr3}, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_KV, type_KV));
+                                                                    hsk, hsv, nh, {nr2, nr3}, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_KV, type_KV, {0, 1, 2, 3}, true, v_is_view_of_k));
                                                        // run fewer test cases permuted
                                                        if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
                                                            test_cases.emplace_back(new test_flash_attn_ext(
-                                                                        hsk, hsv, nh, {nr2, nr3}, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_KV, type_KV, {0, 2, 1, 3}));
+                                                                        hsk, hsv, nh, {nr2, nr3}, kv, nb, mask, sinks, max_bias, logit_softcap, prec, type_KV, type_KV, {0, 2, 1, 3}, true, v_is_view_of_k));
                                                        }
                                                    }
                                                }
@@ -9942,11 +9953,16 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1},  1025,  64, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 2, 1, 3}));
    test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 16384,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));

-    // MLA shape (V is a view of K) with quantized KV
-    // (the test harness builds V as a view of K for this shape; see build_graph)
-    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1},  113,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
-    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1}, 1024,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
-    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1}, 1024,  64, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
+    // MLA shape: the V cache is a sub-view of the K cache, with quantized KV
+    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1},  113,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1}, 1024,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1}, 1024,  64, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
+
+    // more V-is-sub-view-of-K cases: other head shapes, and full views with equal head sizes
+    test_cases.emplace_back(new test_flash_attn_ext(320, 256, 1, {32, 1}, 512, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(192, 128, 4, {8, 1},  512, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(128, 128, 8, {4, 1},  512, 8, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(64,  64,  4, {1, 1},  512, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));

    // large-KV F16 cases (Qwen3.6-27B geometry and a llama-class control): the upstream matrix
    // stops at kv=1024, blind to long-context FA bugs (e.g. the oneDNN SDPA ordering race on BMG).
@@ -239,31 +239,44 @@ Routes → Components → Hooks → Stores → Services → Storage/API

 ### High-Level Architecture

-See: [`docs/architecture/high-level-architecture-simplified.md`](docs/architecture/high-level-architecture-simplified.md)
-
 ```mermaid
 flowchart TB
    subgraph Routes["📍 Routes"]
        R1["/ (Welcome)"]
        R2["/chat/[id]"]
+        R3["/mcp-servers"]
+        R4["/search"]
+        R5["/settings"]
        RL["+layout.svelte"]
    end

    subgraph Components["🧩 Components"]
-        C_Sidebar["ChatSidebar"]
        C_Screen["ChatScreen"]
        C_Form["ChatForm"]
        C_Messages["ChatMessages"]
-        C_ModelsSelector["ModelsSelector"]
+        C_Sidebar["ChatSidebar"]
+        C_Models["ModelsSelector"]
        C_Settings["ChatSettings"]
+        C_Mcp["McpServers"]
+    end
+
+    subgraph Hooks["🔌 Hooks"]
+        H1["use-chat-screen-active-model"]
+        H2["use-processing-state"]
+        H3["use-context-gauge"]
+        H4["use-models-selector"]
+        H5["use-tools-panel"]
    end

    subgraph Stores["🗄️ Stores"]
        S1["chatStore"]
        S2["conversationsStore"]
        S3["modelsStore"]
-        S4["serverStore"]
-        S5["settingsStore"]
+        S4["mcpStore"]
+        S5["agenticStore"]
+        S6["serverStore"]
+        S7["settingsStore"]
+        S8["toolsStore"]
    end

    subgraph Services["⚙️ Services"]
@@ -271,6 +284,9 @@ flowchart TB
        SV2["ModelsService"]
        SV3["PropsService"]
        SV4["DatabaseService"]
+        SV5["MCPService"]
+        SV6["ToolsService"]
+        SV7["SandboxService"]
    end

    subgraph Storage["💾 Storage"]
@@ -282,19 +298,28 @@ flowchart TB
        API1["/v1/chat/completions"]
        API2["/props"]
        API3["/models/*"]
+        API4["/tools"]
    end

    R1 & R2 --> C_Screen
    RL --> C_Sidebar
    C_Screen --> C_Form & C_Messages & C_Settings
-    C_Screen --> S1 & S2
-    C_ModelsSelector --> S3 & S4
+    C_Screen --> H1 & H2 & H3
+    C_Models --> H4
+    C_Mcp --> S4
+    C_Screen --> S1 & S2 & S3
+    C_Models --> S3
+    H1 --> S3
    S1 --> SV1 & SV4
+    S2 --> SV4
    S3 --> SV2 & SV3
+    S4 --> SV5
+    S5 --> SV1 & SV5 & SV6 & SV7
    SV4 --> ST1
    SV1 --> API1
    SV2 --> API3
    SV3 --> API2
+    SV6 --> API4
 ```

 ### Layer Breakdown
@@ -303,6 +328,9 @@ flowchart TB

 - **`/`** - Welcome screen, creates new conversation
 - **`/chat/[id]`** - Active chat interface
+- **`/mcp-servers`** - MCP server management
+- **`/search`** - Conversation search
+- **`/settings`** - Settings (optional `[[section]]`)
 - **`+layout.svelte`** - Sidebar, navigation, global initialization

 #### Components (`src/lib/components/`)
@@ -348,28 +376,68 @@ Components are organized in `app/` (application-specific) and `ui/` (shadcn-svel

 #### Hooks (`src/lib/hooks/`)

- **`useModelChangeValidation`** - Validates model switch against conversation modalities
- **`useProcessingState`** - Tracks streaming progress and token generation
+Hooks are the thin view-layer between components and stores: they own UI concerns (scroll, drag-and-drop, keyboard shortcuts, pickers, selection) and translate store state into view state.
+
+| Hook                            | Responsibility                                                 |
+| ------------------------------- | -------------------------------------------------------------- |
+| `use-chat-screen-active-model`  | Active model resolution + modality capability detection        |
+| `use-processing-state`          | View over `chatStore.processing` for streaming progress/tokens |
+| `use-context-gauge`             | View over `contextStatsStore` for the context usage gauge      |
+| `use-models-selector`           | Model selector dropdown state (loaded/available groups)        |
+| `use-tools-panel`               | Tools panel state                                              |
+| `use-reasoning-menu`            | Reasoning-effort menu state                                    |
+| `use-attachment-menu`           | Attachment menu + modality flags                               |
+| `use-draft-messages`            | Per-chat draft message/files persistence                       |
+| `use-chat-form-pickers`         | Chat form pickers (commands, mentions)                         |
+| `use-debounced-search`          | Shared debounced async search for pickers                      |
+| `use-picker-navigation`         | Picker keyboard navigation                                     |
+| `use-chat-message-edit-context` | Message edit context (content + extras)                        |
+| `use-chat-screen-drag-and-drop` | Drag-and-drop state machine                                    |
+| `use-chat-screen-file-upload`   | File upload queue + capability validation                      |
+| `use-chat-screen-scroll`        | Scroll container binding + navigation guard                    |
+| `use-auto-scroll`               | Auto-scroll controller for streaming                           |
+| `use-marquee-selection`         | Shift+click / marquee range selection                          |
+| `use-keyboard-shortcuts`        | Global keyboard shortcuts                                      |
+| `use-settings-navigation`       | Settings section navigation                                    |
+| `use-pwa`                       | PWA install/update + version mismatch detection                |

 #### Stores (`src/lib/stores/`)

-| Store                | Responsibility                                            |
-| -------------------- | --------------------------------------------------------- |
-| `chatStore`          | Message sending, streaming, abort control, error handling |
-| `conversationsStore` | CRUD for conversations, message branching, navigation     |
-| `modelsStore`        | Model list, selection, loading/unloading (ROUTER)         |
-| `serverStore`        | Server properties, role detection, modalities             |
-| `settingsStore`      | User preferences, parameter sync with server defaults     |
+Stores own reactive application state as Svelte 5 runes. Larger stores are split into directories and compose focused sub-stores behind a narrow host interface (see Architectural Patterns).
+
+| Store                | Responsibility                                                                                                  |
+| -------------------- | --------------------------------------------------------------------------------------------------------------- |
+| `chatStore`          | Chat lifecycle, streaming, abort control, error handling; composes `processing`, `activity`, `streams`, `flows` |
+| `conversationsStore` | Conversation CRUD, message branching, navigation, import/export; composes `preferences`                         |
+| `modelsStore`        | Model list, selection, loading/unloading (ROUTER); composes `props`, `status`                                   |
+| `mcpStore`           | MCP host role: multi-server lifecycle, tool routing; composes `health`, `resources`                             |
+| `agenticStore`       | Multi-turn agentic loop orchestration, tool execution; composes `gates`                                         |
+| `serverStore`        | Server connection state, `/props`, role detection, modalities                                                   |
+| `settingsStore`      | User preferences, theme, parameter sync with server defaults                                                    |
+| `toolsStore`         | Tool registry: server + MCP tools, enabled set for the LLM                                                      |
+| `permissionsStore`   | Persisted tool permission grants                                                                                |
+| `contextStatsStore`  | Context window usage for the active conversation                                                                |
+| `draftMessagesStore` | Per-chat draft message/files                                                                                    |
+| `deviceStore`        | Browser environment signals (mobile, OS, theme)                                                                 |
+| `versionStore`       | Build version information                                                                                       |

 #### Services (`src/lib/services/`)

-| Service                | Responsibility                                  |
-| ---------------------- | ----------------------------------------------- |
-| `ChatService`          | API calls to`/v1/chat/completions`, SSE parsing |
-| `ModelsService`        | `/models`, `/models/load`, `/models/unload`     |
-| `PropsService`         | `/props`, `/props?model=`                       |
-| `DatabaseService`      | IndexedDB operations via Dexie                  |
-| `ParameterSyncService` | Syncs settings with server defaults             |
+Services are a stateless protocol layer: static methods, pure I/O, no reactive state. Stores consume them for all API and storage access.
+
+| Service                       | Responsibility                                                            |
+| ----------------------------- | ------------------------------------------------------------------------- |
+| `ChatService`                 | `/v1/chat/completions` streaming + SSE parsing, message format conversion |
+| `ModelsService`               | `/models`, `/models/load`, `/models/unload`                               |
+| `PropsService`                | `/props`, `/props?model=`                                                 |
+| `DatabaseService`             | IndexedDB operations via Dexie                                            |
+| `MCPService`                  | MCP protocol: transports, connect, list/execute tools, prompts, resources |
+| `ToolsService`                | Server tool list/execute/stream (`/tools`)                                |
+| `SandboxService`              | Browser JS execution in a sandboxed worker                                |
+| `ParameterSyncService`        | Syncs settings with server defaults                                       |
+| `ConversationTransferService` | Conversation import/export JSONL + ZIP format                             |
+| `MigrationService`            | Non-destructive localStorage/IndexedDB migrations                         |
+| `RouterService`               | Dynamic route URL construction                                            |

 ---

@@ -377,8 +445,6 @@ Components are organized in `app/` (application-specific) and `ui/` (shadcn-svel

 ### MODEL Mode (Single Model)

-See: [`docs/flows/data-flow-simplified-model-mode.md`](docs/flows/data-flow-simplified-model-mode.md)
-
 ```mermaid
 sequenceDiagram
    participant User
@@ -388,8 +454,9 @@ sequenceDiagram
    participant API as llama-server

    Note over User,API: Initialization
-    UI->>Stores: initialize()
-    Stores->>DB: load conversations
+    UI->>Stores: initStores() (awaited by route loads)
+    Stores->>Stores: run migrations
+    Stores->>DB: load conversations (background)
    Stores->>API: GET /props
    API-->>Stores: server config
    Stores->>API: GET /v1/models
@@ -408,8 +475,6 @@ sequenceDiagram

 ### ROUTER Mode (Multi-Model)

-See: [`docs/flows/data-flow-simplified-router-mode.md`](docs/flows/data-flow-simplified-router-mode.md)
-
 ```mermaid
 sequenceDiagram
    participant User
@@ -441,17 +506,6 @@ sequenceDiagram
    end
 ```

-### Detailed Flow Diagrams
-
-| Flow          | Description                                | File                                                        |
-| ------------- | ------------------------------------------ | ----------------------------------------------------------- |
-| Chat          | Message lifecycle, streaming, regeneration | [`chat-flow.md`](docs/flows/chat-flow.md)                   |
-| Models        | Loading, unloading, modality caching       | [`models-flow.md`](docs/flows/models-flow.md)               |
-| Server        | Props fetching, role detection             | [`server-flow.md`](docs/flows/server-flow.md)               |
-| Conversations | CRUD, branching, import/export             | [`conversations-flow.md`](docs/flows/conversations-flow.md) |
-| Database      | IndexedDB schema, operations               | [`database-flow.md`](docs/flows/database-flow.md)           |
-| Settings      | Parameter sync, user overrides             | [`settings-flow.md`](docs/flows/settings-flow.md)           |
-
 ---

 ## Architectural Patterns
@@ -505,13 +559,14 @@ Components dispatch actions to stores, stores coordinate with services for I/O,

 ### 3. Per-Conversation State

-Enables concurrent streaming across multiple conversations:
+Enables concurrent streaming across multiple conversations. Loading is tracked
+per conversation by the activity ledger (`chatStore.activity`), while streaming
+state and abort controllers live in per-conversation maps:

 ```typescript
 class ChatStore {
-	chatLoadingStates = new Map<string, boolean>();
-	chatStreamingStates = new Map<string, { response: string; messageId: string }>();
-	abortControllers = new Map<string, AbortController>();
+	chatStreamingStates = new SvelteMap<string, { response: string; messageId: string }>();
+	abortControllers = new SvelteMap<string, AbortController>();
 }
 ```

@@ -567,20 +622,14 @@ get isRouterMode() {

 ### 7. Modality Validation

-Prevents sending attachments to incompatible models:
+Prevents sending attachments to incompatible models. The
+`use-chat-screen-active-model` hook derives the active model's capabilities
+from `modelsStore.props`:

 ```typescript
-// useModelChangeValidation hook
-const validate = (modelId: string) => {
-	const modelModalities = modelsStore.getModelModalities(modelId);
-	const conversationModalities = conversationsStore.usedModalities;
-
-	// Check if model supports all used modalities
-	if (conversationModalities.hasImages && !modelModalities.vision) {
-		return { valid: false, reason: 'Model does not support images' };
-	}
-	// ...
-};
+// use-chat-screen-active-model hook
+const hasVisionModality = $derived.by(() => modelsStore.props.modelSupportsVision(activeModelId));
+const hasAudioModality = $derived.by(() => modelsStore.props.modelSupportsAudio(activeModelId));
 ```

 ### 8. Persistent Storage Strategy
@@ -673,9 +722,6 @@ tools/ui/
 │   └── styles/           # Global styles
 ├── static/               # Static assets
 ├── tests/                # Test files
-├── docs/                 # Architecture diagrams
-│   ├── architecture/     # High-level architecture
-│   └── flows/            # Feature-specific flows
 └── .storybook/           # Storybook configuration
 ```

@@ -1,145 +0,0 @@
-```mermaid
-flowchart TB
-    subgraph Routes["📍 Routes"]
-        R1["/ (Welcome)"]
-        R2["/chat/[id]"]
-        RL["+layout.svelte"]
-    end
-
-    subgraph Components["🧩 Components"]
-        C_Sidebar["ChatSidebar"]
-        C_Screen["ChatScreen"]
-        C_Form["ChatForm"]
-        C_Messages["ChatMessages"]
-        C_Message["ChatMessage"]
-        C_ChatMessageAgenticContent["ChatMessageAgenticContent"]
-        C_MessageEditForm["ChatMessageEditForm"]
-        C_ModelsSelector["ModelsSelector"]
-        C_Settings["ChatSettings"]
-        C_McpSettings["McpServersSettings"]
-        C_McpResourceBrowser["McpResourceBrowser"]
-        C_McpServersSelector["McpServersSelector"]
-    end
-
-    subgraph Hooks["🪝 Hooks"]
-        H1["useModelChangeValidation"]
-        H2["useProcessingState"]
-    end
-
-    subgraph Stores["🗄️ Stores"]
-        S1["chatStore<br/><i>Chat interactions & streaming</i>"]
-        SA["agenticStore<br/><i>Multi-turn agentic loop orchestration</i>"]
-        S2["conversationsStore<br/><i>Conversation data, messages & MCP overrides</i>"]
-        S3["modelsStore<br/><i>Model selection & loading</i>"]
-        S4["serverStore<br/><i>Server props & role detection</i>"]
-        S5["settingsStore<br/><i>User configuration incl. MCP</i>"]
-        S6["mcpStore<br/><i>MCP servers, tools, prompts</i>"]
-        S7["mcpResourceStore<br/><i>MCP resources & attachments</i>"]
-    end
-
-    subgraph Services["⚙️ Services"]
-        SV1["ChatService"]
-        SV2["ModelsService"]
-        SV3["PropsService"]
-        SV4["DatabaseService"]
-        SV5["ParameterSyncService"]
-        SV6["MCPService<br/><i>protocol operations</i>"]
-    end
-
-    subgraph Storage["💾 Storage"]
-        ST1["IndexedDB<br/><i>conversations, messages</i>"]
-        ST2["LocalStorage<br/><i>config, userOverrides, mcpServers</i>"]
-    end
-
-    subgraph APIs["🌐 llama-server API"]
-        API1["/v1/chat/completions"]
-        API2["/props"]
-        API3["/models/*"]
-        API4["/v1/models"]
-    end
-
-    subgraph ExternalMCP["🔌 External MCP Servers"]
-        EXT1["MCP Server 1<br/><i>WebSocket/HTTP/SSE</i>"]
-        EXT2["MCP Server N"]
-    end
-
-    %% Routes → Components
-    R1 & R2 --> C_Screen
-    RL --> C_Sidebar
-
-    %% Layout runs MCP health checks
-    RL --> S6
-
-    %% Component hierarchy
-    C_Screen --> C_Form & C_Messages & C_Settings
-    C_Messages --> C_Message
-    C_Message --> C_ChatMessageAgenticContent
-    C_Message --> C_MessageEditForm
-    C_Form & C_MessageEditForm --> C_ModelsSelector
-    C_Form --> C_McpServersSelector
-    C_Settings --> C_McpSettings
-    C_McpSettings --> C_McpResourceBrowser
-
-    %% Components → Hooks → Stores
-    C_Form & C_Messages --> H1 & H2
-    H1 --> S3 & S4
-    H2 --> S1 & S5
-
-    %% Components → Stores
-    C_Screen --> S1 & S2
-    C_Sidebar --> S2
-    C_ModelsSelector --> S3 & S4
-    C_Settings --> S5
-    C_McpSettings --> S6
-    C_McpResourceBrowser --> S6 & S7
-    C_McpServersSelector --> S6
-    C_Form --> S6
-
-    %% chatStore → agenticStore → mcpStore (agentic loop)
-    S1 --> SA
-    SA --> SV1
-    SA --> S6
-
-    %% Stores → Services
-    S1 --> SV1 & SV4
-    S2 --> SV4
-    S3 --> SV2 & SV3
-    S4 --> SV3
-    S5 --> SV5
-    S6 --> SV6
-    S7 --> SV6
-
-    %% Services → Storage
-    SV4 --> ST1
-    SV5 --> ST2
-
-    %% Services → APIs
-    SV1 --> API1
-    SV2 --> API3 & API4
-    SV3 --> API2
-
-    %% MCP → External Servers
-    SV6 --> EXT1 & EXT2
-
-    %% Styling
-    classDef routeStyle fill:#e1f5fe,stroke:#01579b,stroke-width:2px
-    classDef componentStyle fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
-    classDef hookStyle fill:#fff8e1,stroke:#ff8f00,stroke-width:2px
-    classDef storeStyle fill:#fff3e0,stroke:#e65100,stroke-width:2px
-    classDef serviceStyle fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px
-    classDef storageStyle fill:#fce4ec,stroke:#c2185b,stroke-width:2px
-    classDef apiStyle fill:#e3f2fd,stroke:#1565c0,stroke-width:2px
-    classDef mcpStyle fill:#e0f2f1,stroke:#00695c,stroke-width:2px
-    classDef agenticStyle fill:#e8eaf6,stroke:#283593,stroke-width:2px
-    classDef externalStyle fill:#f3e5f5,stroke:#6a1b9a,stroke-width:2px,stroke-dasharray: 5 5
-
-    class R1,R2,RL routeStyle
-    class C_Sidebar,C_Screen,C_Form,C_Messages,C_Message,C_ChatMessageAgenticContent,C_MessageEditForm,C_ModelsSelector,C_Settings componentStyle
-    class C_McpSettings,C_McpResourceBrowser,C_McpServersSelector componentStyle
-    class H1,H2 hookStyle
-    class S1,S2,S3,S4,S5,SA,S6,S7 storeStyle
-    class SV1,SV2,SV3,SV4,SV5,SV6 serviceStyle
-    class ST1,ST2 storageStyle
-    class API1,API2,API3,API4 apiStyle
-    class EXT1,EXT2 externalStyle
-```
@@ -1,373 +0,0 @@
-```mermaid
-flowchart TB
-subgraph Routes["📍 Routes"]
-R1["/ (+page.svelte)"]
-R2["/chat/[id]"]
-RL["+layout.svelte"]
-end
-
-    subgraph Components["🧩 Components"]
-        direction TB
-        subgraph LayoutComponents["Layout"]
-            C_Sidebar["ChatSidebar"]
-            C_Screen["ChatScreen"]
-        end
-        subgraph ChatUIComponents["Chat UI"]
-            C_Form["ChatForm"]
-            C_Messages["ChatMessages"]
-            C_Message["ChatMessage"]
-            C_MessageUser["ChatMessageUser"]
-            C_MessageEditForm["ChatMessageEditForm"]
-            C_Attach["ChatAttachments"]
-            C_ModelsSelector["ModelsSelector"]
-            C_Settings["ChatSettings"]
-        end
-        subgraph MCPComponents["MCP UI"]
-            C_McpSettings["McpServersSettings"]
-            C_McpServerCard["McpServerCard"]
-            C_McpResourceBrowser["McpResourceBrowser"]
-            C_McpResourcePreview["McpResourcePreview"]
-            C_McpServersSelector["McpServersSelector"]
-        end
-    end
-
-    subgraph Hooks["🪝 Hooks"]
-        H1["useModelChangeValidation"]
-        H2["useProcessingState"]
-        H3["isMobile"]
-    end
-
-    subgraph Stores["🗄️ Stores"]
-        direction TB
-        subgraph S1["chatStore"]
-            S1State["<b>State:</b><br/>isLoading, currentResponse<br/>errorDialogState<br/>activeProcessingState<br/>chatLoadingStates<br/>chatStreamingStates<br/>abortControllers<br/>processingStates<br/>activeConversationId<br/>isStreamingActive"]
-            S1LoadState["<b>Loading State:</b><br/>setChatLoading()<br/>isChatLoading()<br/>syncLoadingStateForChat()<br/>clearUIState()<br/>isChatLoadingPublic()<br/>getAllLoadingChats()<br/>getAllStreamingChats()"]
-            S1ProcState["<b>Processing State:</b><br/>setActiveProcessingConversation()<br/>getProcessingState()<br/>clearProcessingState()<br/>getActiveProcessingState()<br/>updateProcessingStateFromTimings()<br/>getCurrentProcessingStateSync()<br/>restoreProcessingStateFromMessages()"]
-            S1Stream["<b>Streaming:</b><br/>streamChatCompletion()<br/>startStreaming()<br/>stopStreaming()<br/>stopGeneration()<br/>isStreaming()"]
-            S1Error["<b>Error Handling:</b><br/>showErrorDialog()<br/>dismissErrorDialog()<br/>isAbortError()"]
-            S1Msg["<b>Message Operations:</b><br/>addMessage()<br/>sendMessage()<br/>updateMessage()<br/>deleteMessage()<br/>getDeletionInfo()"]
-            S1Regen["<b>Regeneration:</b><br/>regenerateMessage()<br/>regenerateMessageWithBranching()<br/>continueAssistantMessage()"]
-            S1Edit["<b>Editing:</b><br/>editAssistantMessage()<br/>editUserMessagePreserveResponses()<br/>editMessageWithBranching()<br/>clearEditMode()<br/>isEditModeActive()<br/>getAddFilesHandler()<br/>setEditModeActive()"]
-            S1Utils["<b>Utilities:</b><br/>getApiOptions()<br/>parseTimingData()<br/>getOrCreateAbortController()<br/>getConversationModel()"]
-        end
-        subgraph SA["agenticStore"]
-            SAState["<b>State:</b><br/>sessions (Map)<br/>isAnyRunning"]
-            SASession["<b>Session Management:</b><br/>getSession()<br/>updateSession()<br/>clearSession()<br/>getActiveSessions()<br/>isRunning()<br/>currentTurn()<br/>totalToolCalls()<br/>lastError()<br/>streamingToolCall()"]
-            SAConfig["<b>Configuration:</b><br/>getConfig()<br/>maxTurns, maxToolPreviewLines"]
-            SAFlow["<b>Agentic Loop:</b><br/>runAgenticFlow()<br/>executeAgenticLoop()<br/>normalizeToolCalls()<br/>emitToolCallResult()<br/>extractBase64Attachments()"]
-        end
-        subgraph S2["conversationsStore"]
-            S2State["<b>State:</b><br/>conversations<br/>activeConversation<br/>activeMessages<br/>isInitialized<br/>pendingMcpServerOverrides<br/>titleUpdateConfirmationCallback"]
-            S2Lifecycle["<b>Lifecycle:</b><br/>initialize()<br/>loadConversations()<br/>clearActiveConversation()"]
-            S2ConvCRUD["<b>Conversation CRUD:</b><br/>createConversation()<br/>loadConversation()<br/>deleteConversation()<br/>deleteAll()<br/>updateConversationName()<br/>updateConversationTitleWithConfirmation()"]
-            S2MsgMgmt["<b>Message Management:</b><br/>refreshActiveMessages()<br/>addMessageToActive()<br/>updateMessageAtIndex()<br/>findMessageIndex()<br/>sliceActiveMessages()<br/>removeMessageAtIndex()<br/>getConversationMessages()"]
-            S2Nav["<b>Navigation:</b><br/>navigateToSibling()<br/>updateCurrentNode()<br/>updateConversationTimestamp()"]
-            S2McpOverrides["<b>MCP Per-Chat Overrides:</b><br/>getMcpServerOverride()<br/>getAllMcpServerOverrides()<br/>setMcpServerOverride()<br/>toggleMcpServerForChat()<br/>removeMcpServerOverride()<br/>isMcpServerEnabledForChat()<br/>clearPendingMcpServerOverrides()"]
-            S2Export["<b>Import/Export:</b><br/>downloadConversation()<br/>exportAllConversations()<br/>importConversations()<br/>importConversationsData()<br/>triggerDownload()"]
-            S2Utils["<b>Utilities:</b><br/>setTitleUpdateConfirmationCallback()"]
-        end
-        subgraph S3["modelsStore"]
-            S3State["<b>State:</b><br/>models, routerModels<br/>selectedModelId<br/>selectedModelName<br/>loading, updating, error<br/>modelLoadingStates<br/>modelPropsCache<br/>modelPropsFetching<br/>propsCacheVersion"]
-            S3Getters["<b>Computed Getters:</b><br/>selectedModel<br/>loadedModelIds<br/>loadingModelIds<br/>singleModelName"]
-            S3Modal["<b>Modalities:</b><br/>getModelModalities()<br/>modelSupportsVision()<br/>modelSupportsAudio()<br/>getModelModalitiesArray()<br/>getModelProps()<br/>updateModelModalities()"]
-            S3Status["<b>Status Queries:</b><br/>isModelLoaded()<br/>isModelOperationInProgress()<br/>getModelStatus()<br/>isModelPropsFetching()"]
-            S3Fetch["<b>Data Fetching:</b><br/>fetch()<br/>fetchRouterModels()<br/>fetchModelProps()<br/>fetchModalitiesForLoadedModels()"]
-            S3Select["<b>Model Selection:</b><br/>selectModelById()<br/>selectModelByName()<br/>clearSelection()<br/>findModelByName()<br/>findModelById()<br/>hasModel()"]
-            S3LoadUnload["<b>Loading/Unloading Models:</b><br/>loadModel()<br/>unloadModel()<br/>ensureModelLoaded()<br/>waitForModelStatus()<br/>pollForModelStatus()"]
-            S3Utils["<b>Utilities:</b><br/>toDisplayName()<br/>clear()"]
-        end
-        subgraph S4["serverStore"]
-            S4State["<b>State:</b><br/>props<br/>loading, error<br/>role<br/>fetchPromise"]
-            S4Getters["<b>Getters:</b><br/>defaultParams<br/>contextSize<br/>isRouterMode<br/>isModelMode"]
-            S4Data["<b>Data Handling:</b><br/>fetch()<br/>getErrorMessage()<br/>clear()"]
-            S4Utils["<b>Utilities:</b><br/>detectRole()"]
-        end
-        subgraph S5["settingsStore"]
-            S5State["<b>State:</b><br/>config<br/>theme<br/>isInitialized<br/>userOverrides"]
-            S5Lifecycle["<b>Lifecycle:</b><br/>initialize()<br/>loadConfig()<br/>saveConfig()<br/>loadTheme()<br/>saveTheme()"]
-            S5Update["<b>Config Updates:</b><br/>updateConfig()<br/>updateMultipleConfig()<br/>updateTheme()"]
-            S5Reset["<b>Reset:</b><br/>resetConfig()<br/>resetTheme()<br/>resetAll()<br/>resetParameterToServerDefault()"]
-            S5Sync["<b>Server Sync:</b><br/>syncWithServerDefaults()<br/>forceSyncWithServerDefaults()"]
-            S5Utils["<b>Utilities:</b><br/>getConfig()<br/>getAllConfig()<br/>getParameterInfo()<br/>getParameterDiff()<br/>getServerDefaults()<br/>clearAllUserOverrides()"]
-        end
-        subgraph S6["mcpStore"]
-            S6State["<b>State:</b><br/>isInitializing, error<br/>toolCount, connectedServers<br/>healthChecks (Map)<br/>connections (Map)<br/>toolsIndex (Map)"]
-            S6Lifecycle["<b>Lifecycle:</b><br/>ensureInitialized()<br/>initialize()<br/>shutdown()<br/>acquireConnection()<br/>releaseConnection()"]
-            S6Health["<b>Health Checks:</b><br/>runHealthCheck()<br/>runHealthChecksForServers()<br/>updateHealthCheck()<br/>getHealthCheckState()<br/>clearHealthCheck()"]
-            S6Servers["<b>Server Management:</b><br/>getServers()<br/>addServer()<br/>updateServer()<br/>removeServer()<br/>getServerById()<br/>getServerDisplayName()"]
-            S6Tools["<b>Tool Operations:</b><br/>getToolDefinitionsForLLM()<br/>getToolNames()<br/>hasTool()<br/>getToolServer()<br/>executeTool()<br/>executeToolByName()"]
-            S6Prompts["<b>Prompt Operations:</b><br/>getAllPrompts()<br/>getPrompt()<br/>hasPromptsCapability()<br/>getPromptCompletions()"]
-        end
-        subgraph S7["mcpResourceStore"]
-            S7State["<b>State:</b><br/>serverResources (Map)<br/>cachedResources (Map)<br/>subscriptions (Map)<br/>attachments[]<br/>isLoading"]
-            S7Resources["<b>Resource Discovery:</b><br/>setServerResources()<br/>getServerResources()<br/>getAllResourceInfos()<br/>getAllTemplateInfos()<br/>clearServerResources()"]
-            S7Cache["<b>Caching:</b><br/>cacheResourceContent()<br/>getCachedContent()<br/>invalidateCache()<br/>clearCache()"]
-            S7Subs["<b>Subscriptions:</b><br/>addSubscription()<br/>removeSubscription()<br/>isSubscribed()<br/>handleResourceUpdate()"]
-            S7Attach["<b>Attachments:</b><br/>addAttachment()<br/>updateAttachmentContent()<br/>removeAttachment()<br/>clearAttachments()<br/>toMessageExtras()"]
-        end
-
-        subgraph ReactiveExports["⚡ Reactive Exports"]
-            direction LR
-            subgraph ChatExports["chatStore"]
-                RE1["isLoading()"]
-                RE2["currentResponse()"]
-                RE3["errorDialog()"]
-                RE4["activeProcessingState()"]
-                RE5["isChatStreaming()"]
-                RE6["isChatLoading()"]
-                RE7["getChatStreaming()"]
-                RE8["getAllLoadingChats()"]
-                RE9["getAllStreamingChats()"]
-                RE9a["isEditModeActive()"]
-                RE9b["getAddFilesHandler()"]
-                RE9c["setEditModeActive()"]
-                RE9d["clearEditMode()"]
-            end
-            subgraph AgenticExports["agenticStore"]
-                REA1["agenticIsRunning()"]
-                REA2["agenticCurrentTurn()"]
-                REA3["agenticTotalToolCalls()"]
-                REA4["agenticLastError()"]
-                REA5["agenticStreamingToolCall()"]
-                REA6["agenticIsAnyRunning()"]
-            end
-            subgraph ConvExports["conversationsStore"]
-                RE10["conversations()"]
-                RE11["activeConversation()"]
-                RE12["activeMessages()"]
-                RE13["isConversationsInitialized()"]
-            end
-            subgraph ModelsExports["modelsStore"]
-                RE15["modelOptions()"]
-                RE16["routerModels()"]
-                RE17["modelsLoading()"]
-                RE18["modelsUpdating()"]
-                RE19["modelsError()"]
-                RE20["selectedModelId()"]
-                RE21["selectedModelName()"]
-                RE22["selectedModelOption()"]
-                RE23["loadedModelIds()"]
-                RE24["loadingModelIds()"]
-                RE25["propsCacheVersion()"]
-                RE26["singleModelName()"]
-            end
-            subgraph ServerExports["serverStore"]
-                RE27["serverProps()"]
-                RE28["serverLoading()"]
-                RE29["serverError()"]
-                RE30["serverRole()"]
-                RE31["defaultParams()"]
-                RE32["contextSize()"]
-                RE33["isRouterMode()"]
-                RE34["isModelMode()"]
-            end
-            subgraph SettingsExports["settingsStore"]
-                RE35["config()"]
-                RE36["theme()"]
-                RE37["isInitialized()"]
-            end
-            subgraph MCPExports["mcpStore / mcpResourceStore"]
-                RE38["mcpResources()"]
-                RE39["mcpResourceAttachments()"]
-                RE40["mcpHasResourceAttachments()"]
-                RE41["mcpTotalResourceCount()"]
-                RE42["mcpResourcesLoading()"]
-            end
-        end
-    end
-
-    subgraph Services["⚙️ Services"]
-        direction TB
-        subgraph SV1["ChatService"]
-            SV1Msg["<b>Messaging:</b><br/>sendMessage()"]
-            SV1Stream["<b>Streaming:</b><br/>handleStreamResponse()<br/>handleNonStreamResponse()"]
-            SV1Convert["<b>Conversion:</b><br/>convertDbMessageToApiChatMessageData()<br/>mergeToolCallDeltas()"]
-            SV1Utils["<b>Utilities:</b><br/>stripReasoningContent()<br/>extractModelName()<br/>parseErrorResponse()"]
-        end
-        subgraph SV2["ModelsService"]
-            SV2List["<b>Listing:</b><br/>list()<br/>listRouter()"]
-            SV2LoadUnload["<b>Load/Unload:</b><br/>load()<br/>unload()"]
-            SV2Status["<b>Status:</b><br/>isModelLoaded()<br/>isModelLoading()"]
-        end
-        subgraph SV3["PropsService"]
-            SV3Fetch["<b>Fetching:</b><br/>fetch()<br/>fetchForModel()"]
-        end
-        subgraph SV4["DatabaseService"]
-            SV4Conv["<b>Conversations:</b><br/>createConversation()<br/>getConversation()<br/>getAllConversations()<br/>updateConversation()<br/>deleteConversation()"]
-            SV4Msg["<b>Messages:</b><br/>createMessageBranch()<br/>createRootMessage()<br/>createSystemMessage()<br/>getConversationMessages()<br/>updateMessage()<br/>deleteMessage()<br/>deleteMessageCascading()"]
-            SV4Node["<b>Navigation:</b><br/>updateCurrentNode()"]
-            SV4Import["<b>Import:</b><br/>importConversations()"]
-        end
-        subgraph SV5["ParameterSyncService"]
-            SV5Extract["<b>Extraction:</b><br/>extractServerDefaults()"]
-            SV5Merge["<b>Merging:</b><br/>mergeWithServerDefaults()"]
-            SV5Info["<b>Info:</b><br/>getParameterInfo()<br/>canSyncParameter()<br/>getSyncableParameterKeys()<br/>validateServerParameter()"]
-            SV5Diff["<b>Diff:</b><br/>createParameterDiff()"]
-        end
-        subgraph SV6["MCPService"]
-            SV6Transport["<b>Transport:</b><br/>createTransport()<br/>WebSocket / StreamableHTTP / SSE"]
-            SV6Conn["<b>Connection:</b><br/>connect()<br/>disconnect()"]
-            SV6Tools["<b>Tools:</b><br/>listTools()<br/>callTool()"]
-            SV6Prompts["<b>Prompts:</b><br/>listPrompts()<br/>getPrompt()"]
-            SV6Resources["<b>Resources:</b><br/>listResources()<br/>listResourceTemplates()<br/>readResource()<br/>subscribeResource()<br/>unsubscribeResource()"]
-            SV6Complete["<b>Completions:</b><br/>complete()"]
-        end
-    end
-
-    subgraph ExternalMCP["🔌 External MCP Servers"]
-        EXT1["MCP Server 1<br/>(WebSocket/StreamableHTTP/SSE)"]
-        EXT2["MCP Server N"]
-    end
-
-    subgraph Storage["💾 Storage"]
-        ST1["IndexedDB"]
-        ST2["conversations"]
-        ST3["messages"]
-        ST5["LocalStorage"]
-        ST6["config"]
-        ST7["userOverrides"]
-        ST8["mcpServers"]
-    end
-
-    subgraph APIs["🌐 llama-server API"]
-        API1["/v1/chat/completions"]
-        API2["/props<br/>/props?model="]
-        API3["/models<br/>/models/load<br/>/models/unload"]
-        API4["/v1/models"]
-    end
-
-    %% Routes render Components
-    R1 --> C_Screen
-    R2 --> C_Screen
-    RL --> C_Sidebar
-
-    %% Layout runs MCP health checks on startup
-    RL --> S6
-
-    %% Component hierarchy
-    C_Screen --> C_Form & C_Messages & C_Settings
-    C_Messages --> C_Message
-    C_Message --> C_MessageUser
-    C_MessageUser --> C_MessageEditForm
-    C_MessageEditForm --> C_ModelsSelector
-    C_MessageEditForm --> C_Attach
-    C_Form --> C_ModelsSelector
-    C_Form --> C_Attach
-    C_Form --> C_McpServersSelector
-    C_Message --> C_Attach
-
-    %% MCP Components hierarchy
-    C_Settings --> C_McpSettings
-    C_McpSettings --> C_McpServerCard
-    C_McpServerCard --> C_McpResourceBrowser
-    C_McpResourceBrowser --> C_McpResourcePreview
-
-    %% Components use Hooks
-    C_Form --> H1
-    C_Message --> H1 & H2
-    C_MessageEditForm --> H1
-    C_Screen --> H2
-
-    %% Hooks use Stores
-    H1 --> S3 & S4
-    H2 --> S1 & S5
-
-    %% Components use Stores
-    C_Screen --> S1 & S2
-    C_Messages --> S2
-    C_Message --> S1 & S2 & S3
-    C_Form --> S1 & S3 & S6
-    C_Sidebar --> S2
-    C_ModelsSelector --> S3 & S4
-    C_Settings --> S5
-    C_McpSettings --> S6
-    C_McpServerCard --> S6
-    C_McpResourceBrowser --> S6 & S7
-    C_McpServersSelector --> S6
-
-    %% Stores export Reactive State
-    S1 -. exports .-> ChatExports
-    SA -. exports .-> AgenticExports
-    S2 -. exports .-> ConvExports
-    S3 -. exports .-> ModelsExports
-    S4 -. exports .-> ServerExports
-    S5 -. exports .-> SettingsExports
-    S6 -. exports .-> MCPExports
-    S7 -. exports .-> MCPExports
-
-    %% chatStore → agenticStore (agentic loop orchestration)
-    S1 --> SA
-    SA --> SV1
-    SA --> S6
-
-    %% Stores use Services
-    S1 --> SV1 & SV4
-    S2 --> SV4
-    S3 --> SV2 & SV3
-    S4 --> SV3
-    S5 --> SV5
-    S6 --> SV6
-    S7 --> SV6
-
-    %% Services to Storage
-    SV4 --> ST1
-    ST1 --> ST2 & ST3
-    SV5 --> ST5
-    ST5 --> ST6 & ST7 & ST8
-
-    %% Services to APIs
-    SV1 --> API1
-    SV2 --> API3 & API4
-    SV3 --> API2
-
-    %% MCP → External Servers
-    SV6 --> EXT1 & EXT2
-
-    %% Styling
-    classDef routeStyle fill:#e1f5fe,stroke:#01579b,stroke-width:2px
-    classDef componentStyle fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
-    classDef componentGroupStyle fill:#e1bee7,stroke:#7b1fa2,stroke-width:1px
-    classDef hookStyle fill:#fff8e1,stroke:#ff8f00,stroke-width:2px
-    classDef storeStyle fill:#fff3e0,stroke:#e65100,stroke-width:2px
-    classDef stateStyle fill:#ffe0b2,stroke:#e65100,stroke-width:1px
-    classDef methodStyle fill:#ffecb3,stroke:#e65100,stroke-width:1px
-    classDef reactiveStyle fill:#fffde7,stroke:#f9a825,stroke-width:1px
-    classDef serviceStyle fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px
-    classDef serviceMStyle fill:#c8e6c9,stroke:#2e7d32,stroke-width:1px
-    classDef externalStyle fill:#f3e5f5,stroke:#6a1b9a,stroke-width:2px,stroke-dasharray: 5 5
-    classDef storageStyle fill:#fce4ec,stroke:#c2185b,stroke-width:2px
-    classDef apiStyle fill:#e3f2fd,stroke:#1565c0,stroke-width:2px
-
-    class R1,R2,RL routeStyle
-    class C_Sidebar,C_Screen,C_Form,C_Messages,C_Message,C_MessageUser,C_MessageEditForm componentStyle
-    class C_ModelsSelector,C_Settings componentStyle
-    class C_Attach componentStyle
-    class C_McpSettings,C_McpServerCard,C_McpResourceBrowser,C_McpResourcePreview,C_McpServersSelector componentStyle
-    class H1,H2,H3 hookStyle
-    class LayoutComponents,ChatUIComponents,MCPComponents componentGroupStyle
-    class Hooks hookStyle
-    classDef agenticStyle fill:#e8eaf6,stroke:#283593,stroke-width:2px
-    classDef agenticMethodStyle fill:#c5cae9,stroke:#283593,stroke-width:1px
-
-    class S1,S2,S3,S4,S5,SA,S6,S7 storeStyle
-    class S1State,S2State,S3State,S4State,S5State,SAState,S6State,S7State stateStyle
-    class S1Msg,S1Regen,S1Edit,S1Stream,S1LoadState,S1ProcState,S1Error,S1Utils methodStyle
-    class SASession,SAConfig,SAFlow methodStyle
-    class S2Lifecycle,S2ConvCRUD,S2MsgMgmt,S2Nav,S2McpOverrides,S2Export,S2Utils methodStyle
-    class S3Getters,S3Modal,S3Status,S3Fetch,S3Select,S3LoadUnload,S3Utils methodStyle
-    class S4Getters,S4Data,S4Utils methodStyle
-    class S5Lifecycle,S5Update,S5Reset,S5Sync,S5Utils methodStyle
-    class S6Lifecycle,S6Health,S6Servers,S6Tools,S6Prompts methodStyle
-    class S7Resources,S7Cache,S7Subs,S7Attach methodStyle
-    class ChatExports,AgenticExports,ConvExports,ModelsExports,ServerExports,SettingsExports,MCPExports reactiveStyle
-    class SV1,SV2,SV3,SV4,SV5,SV6 serviceStyle
-    class SV6Transport,SV6Conn,SV6Tools,SV6Prompts,SV6Resources,SV6Complete serviceMStyle
-    class EXT1,EXT2 externalStyle
-    class SV1Msg,SV1Stream,SV1Convert,SV1Utils serviceMStyle
-    class SV2List,SV2LoadUnload,SV2Status serviceMStyle
-    class SV3Fetch serviceMStyle
-    class SV4Conv,SV4Msg,SV4Node,SV4Import serviceMStyle
-    class SV5Extract,SV5Merge,SV5Info,SV5Diff serviceMStyle
-    class ST1,ST2,ST3,ST5,ST6,ST7,ST8 storageStyle
-    class API1,API2,API3,API4 apiStyle
-```
@@ -1,228 +0,0 @@
-```mermaid
-sequenceDiagram
-    participant UI as 🧩 ChatForm / ChatMessage
-    participant chatStore as 🗄️ chatStore
-    participant agenticStore as 🗄️ agenticStore
-    participant convStore as 🗄️ conversationsStore
-    participant settingsStore as 🗄️ settingsStore
-    participant mcpStore as 🗄️ mcpStore
-    participant ChatSvc as ⚙️ ChatService
-    participant DbSvc as ⚙️ DatabaseService
-    participant API as 🌐 /v1/chat/completions
-
-    Note over chatStore: State:<br/>isLoading, currentResponse<br/>errorDialogState, activeProcessingState<br/>chatLoadingStates (Map)<br/>chatStreamingStates (Map)<br/>abortControllers (Map)<br/>processingStates (Map)
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 💬 SEND MESSAGE
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>chatStore: sendMessage(content, extras)
-    activate chatStore
-
-    chatStore->>chatStore: setChatLoading(convId, true)
-    chatStore->>chatStore: clearChatStreaming(convId)
-
-    alt no active conversation
-        chatStore->>convStore: createConversation()
-        Note over convStore: → see conversations-flow.mmd
-    end
-
-    chatStore->>mcpStore: consumeResourceAttachmentsAsExtras()
-    Note right of mcpStore: Converts pending MCP resource<br/>attachments into message extras
-
-    chatStore->>chatStore: addMessage("user", content, extras)
-    chatStore->>DbSvc: createMessageBranch(userMsg, parentId)
-    chatStore->>convStore: addMessageToActive(userMsg)
-    chatStore->>convStore: updateCurrentNode(userMsg.id)
-
-    chatStore->>chatStore: createAssistantMessage(userMsg.id)
-    chatStore->>DbSvc: createMessageBranch(assistantMsg, userMsg.id)
-    chatStore->>convStore: addMessageToActive(assistantMsg)
-
-    chatStore->>chatStore: streamChatCompletion(messages, assistantMsg)
-    deactivate chatStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 🌊 STREAMING (with agentic flow detection)
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    activate chatStore
-    chatStore->>chatStore: startStreaming()
-    Note right of chatStore: isStreamingActive = true
-
-    chatStore->>chatStore: setActiveProcessingConversation(convId)
-    chatStore->>chatStore: getOrCreateAbortController(convId)
-    Note right of chatStore: abortControllers.set(convId, new AbortController())
-
-    chatStore->>chatStore: getApiOptions()
-    Note right of chatStore: Merge from settingsStore.config:<br/>temperature, max_tokens, top_p, etc.
-
-    alt agenticConfig.enabled && mcpStore has connected servers
-        chatStore->>agenticStore: runAgenticFlow(convId, messages, assistantMsg, options, signal)
-        Note over agenticStore: Multi-turn agentic loop:<br/>1. Call ChatService.sendMessage()<br/>2. If response has tool_calls → execute via mcpStore<br/>3. Append tool results as messages<br/>4. Loop until no more tool_calls or maxTurns<br/>→ see agentic flow details below
-        agenticStore-->>chatStore: final response with timings
-    else standard (non-agentic) flow
-        chatStore->>ChatSvc: sendMessage(messages, options, signal)
-    end
-
-    activate ChatSvc
-
-    ChatSvc->>ChatSvc: convertDbMessageToApiChatMessageData(messages)
-    Note right of ChatSvc: DatabaseMessage[] → ApiChatMessageData[]<br/>Process attachments (images, PDFs, audio)
-
-    ChatSvc->>API: POST /v1/chat/completions
-    Note right of API: {messages, model?, stream: true, ...params}
-
-    loop SSE chunks
-        API-->>ChatSvc: data: {"choices":[{"delta":{...}}]}
-        ChatSvc->>ChatSvc: handleStreamResponse(response)
-
-        alt content chunk
-            ChatSvc-->>chatStore: onChunk(content)
-            chatStore->>chatStore: setChatStreaming(convId, response, msgId)
-            Note right of chatStore: currentResponse = $state(accumulated)
-            chatStore->>convStore: updateMessageAtIndex(idx, {content})
-        end
-
-        alt reasoning chunk
-            ChatSvc-->>chatStore: onReasoningChunk(reasoning)
-            chatStore->>convStore: updateMessageAtIndex(idx, {thinking})
-        end
-
-        alt tool_calls chunk
-            ChatSvc-->>chatStore: onToolCallChunk(toolCalls)
-            chatStore->>convStore: updateMessageAtIndex(idx, {toolCalls})
-        end
-
-        alt model info
-            ChatSvc-->>chatStore: onModel(modelName)
-            chatStore->>chatStore: recordModel(modelName)
-            chatStore->>DbSvc: updateMessage(msgId, {model})
-        end
-
-        alt timings (during stream)
-            ChatSvc-->>chatStore: onTimings(timings, promptProgress)
-            chatStore->>chatStore: updateProcessingStateFromTimings()
-        end
-
-        chatStore-->>UI: reactive $state update
-    end
-
-    API-->>ChatSvc: data: [DONE]
-    ChatSvc-->>chatStore: onComplete(content, reasoning, timings, toolCalls)
-    deactivate ChatSvc
-
-    chatStore->>chatStore: stopStreaming()
-    chatStore->>DbSvc: updateMessage(msgId, {content, timings, model})
-    chatStore->>convStore: updateCurrentNode(msgId)
-    chatStore->>chatStore: setChatLoading(convId, false)
-    chatStore->>chatStore: clearChatStreaming(convId)
-    chatStore->>chatStore: clearProcessingState(convId)
-    deactivate chatStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: ⏹️ STOP GENERATION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>chatStore: stopGeneration()
-    activate chatStore
-    chatStore->>chatStore: savePartialResponseIfNeeded(convId)
-    Note right of chatStore: Save currentResponse to DB if non-empty
-    chatStore->>chatStore: abortControllers.get(convId).abort()
-    Note right of chatStore: fetch throws AbortError → caught by isAbortError()
-    chatStore->>chatStore: stopStreaming()
-    chatStore->>chatStore: setChatLoading(convId, false)
-    chatStore->>chatStore: clearChatStreaming(convId)
-    chatStore->>chatStore: clearProcessingState(convId)
-    deactivate chatStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 🔁 REGENERATE
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>chatStore: regenerateMessageWithBranching(msgId, model?)
-    activate chatStore
-    chatStore->>convStore: findMessageIndex(msgId)
-    chatStore->>chatStore: Get parent of target message
-    chatStore->>chatStore: createAssistantMessage(parentId)
-    chatStore->>DbSvc: createMessageBranch(newAssistantMsg, parentId)
-    chatStore->>convStore: refreshActiveMessages()
-    Note right of chatStore: Same streaming flow
-    chatStore->>chatStore: streamChatCompletion(...)
-    deactivate chatStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: ➡️ CONTINUE
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>chatStore: continueAssistantMessage(msgId)
-    activate chatStore
-    chatStore->>chatStore: Get existing content from message
-    chatStore->>chatStore: streamChatCompletion(..., existingContent)
-    Note right of chatStore: Appends to existing message content
-    deactivate chatStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: ✏️ EDIT USER MESSAGE
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>chatStore: editMessageWithBranching(msgId, newContent, extras)
-    activate chatStore
-    chatStore->>chatStore: Get parent of target message
-    chatStore->>DbSvc: createMessageBranch(editedMsg, parentId)
-    chatStore->>convStore: refreshActiveMessages()
-    Note right of chatStore: Creates new branch, original preserved
-    chatStore->>chatStore: createAssistantMessage(editedMsg.id)
-    chatStore->>chatStore: streamChatCompletion(...)
-    Note right of chatStore: Automatically regenerates response
-    deactivate chatStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: ❌ ERROR HANDLING
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over chatStore: On stream error (non-abort):
-    chatStore->>chatStore: showErrorDialog(type, message)
-    Note right of chatStore: errorDialogState = {type: 'timeout'|'server', message}
-    chatStore->>convStore: removeMessageAtIndex(failedMsgIdx)
-    chatStore->>DbSvc: deleteMessage(failedMsgId)
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 🤖 AGENTIC LOOP (when agenticConfig.enabled)
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over agenticStore: agenticStore.runAgenticFlow(convId, messages, assistantMsg, options, signal)
-    activate agenticStore
-    agenticStore->>agenticStore: getSession(convId) or create new
-    agenticStore->>agenticStore: updateSession(turn: 0, running: true)
-
-    loop executeAgenticLoop (until no tool_calls or maxTurns)
-        agenticStore->>agenticStore: turn++
-        agenticStore->>ChatSvc: sendMessage(messages, options, signal)
-        ChatSvc->>API: POST /v1/chat/completions
-        API-->>ChatSvc: response with potential tool_calls
-        ChatSvc-->>agenticStore: onComplete(content, reasoning, timings, toolCalls)
-
-        alt response has tool_calls
-            agenticStore->>agenticStore: normalizeToolCalls(toolCalls)
-            loop for each tool_call
-                agenticStore->>agenticStore: updateSession(streamingToolCall)
-                agenticStore->>mcpStore: executeTool(mcpCall, signal)
-                mcpStore-->>agenticStore: tool result
-                agenticStore->>agenticStore: extractBase64Attachments(result)
-                agenticStore->>agenticStore: emitToolCallResult(convId, ...)
-                agenticStore->>convStore: addMessageToActive(toolResultMsg)
-                agenticStore->>DbSvc: createMessageBranch(toolResultMsg)
-            end
-            agenticStore->>agenticStore: Create new assistantMsg for next turn
-            Note right of agenticStore: Continue loop with updated messages
-        else no tool_calls (final response)
-            agenticStore->>agenticStore: buildFinalTimings(allTurns)
-            Note right of agenticStore: Break loop, return final response
-        end
-    end
-
-    agenticStore->>agenticStore: updateSession(running: false)
-    agenticStore-->>chatStore: final content, timings, model
-    deactivate agenticStore
-```
@@ -1,183 +0,0 @@
-```mermaid
-sequenceDiagram
-    participant UI as 🧩 ChatSidebar / ChatScreen
-    participant convStore as 🗄️ conversationsStore
-    participant chatStore as 🗄️ chatStore
-    participant DbSvc as ⚙️ DatabaseService
-    participant IDB as 💾 IndexedDB
-
-    Note over convStore: State:<br/>conversations: DatabaseConversation[]<br/>activeConversation: DatabaseConversation | null<br/>activeMessages: DatabaseMessage[]<br/>isInitialized: boolean<br/>pendingMcpServerOverrides: Map&lt;string, McpServerOverride&gt;
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,IDB: 🚀 INITIALIZATION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over convStore: Auto-initialized in constructor (browser only)
-    convStore->>convStore: initialize()
-    activate convStore
-    convStore->>convStore: loadConversations()
-    convStore->>DbSvc: getAllConversations()
-    DbSvc->>IDB: SELECT * FROM conversations ORDER BY lastModified DESC
-    IDB-->>DbSvc: Conversation[]
-    DbSvc-->>convStore: conversations
-    convStore->>convStore: conversations = $state(data)
-    convStore->>convStore: isInitialized = true
-    deactivate convStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,IDB: ➕ CREATE CONVERSATION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>convStore: createConversation(name?)
-    activate convStore
-    convStore->>DbSvc: createConversation(name || "New Chat")
-    DbSvc->>IDB: INSERT INTO conversations
-    IDB-->>DbSvc: conversation {id, name, lastModified, currNode: ""}
-    DbSvc-->>convStore: conversation
-    convStore->>convStore: conversations.unshift(conversation)
-    convStore->>convStore: activeConversation = $state(conversation)
-    convStore->>convStore: activeMessages = $state([])
-
-    alt pendingMcpServerOverrides has entries
-        loop each pending override
-            convStore->>DbSvc: Store MCP server override for new conversation
-        end
-        convStore->>convStore: clearPendingMcpServerOverrides()
-    end
-    deactivate convStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,IDB: 📂 LOAD CONVERSATION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>convStore: loadConversation(convId)
-    activate convStore
-    convStore->>DbSvc: getConversation(convId)
-    DbSvc->>IDB: SELECT * FROM conversations WHERE id = ?
-    IDB-->>DbSvc: conversation
-    convStore->>convStore: activeConversation = $state(conversation)
-
-    convStore->>convStore: refreshActiveMessages()
-    convStore->>DbSvc: getConversationMessages(convId)
-    DbSvc->>IDB: SELECT * FROM messages WHERE convId = ?
-    IDB-->>DbSvc: allMessages[]
-    convStore->>convStore: filterByLeafNodeId(allMessages, currNode)
-    Note right of convStore: Filter to show only current branch path
-    convStore->>convStore: activeMessages = $state(filtered)
-
-    Note right of convStore: Route (+page.svelte) then calls:<br/>chatStore.syncLoadingStateForChat(convId)
-    deactivate convStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,IDB: 🌳 MESSAGE BRANCHING MODEL
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over IDB: Message Tree Structure:<br/>- Each message has parent (null for root)<br/>- Each message has children[] array<br/>- Conversation.currNode points to active leaf<br/>- filterByLeafNodeId() traverses from root to currNode
-
-    rect rgb(240, 240, 255)
-        Note over convStore: Example Branch Structure:
-        Note over convStore: root → user1 → assistant1 → user2 → assistant2a (currNode)<br/>                                    ↘ assistant2b (alt branch)
-    end
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,IDB: ↔️ BRANCH NAVIGATION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>convStore: navigateToSibling(msgId, direction)
-    activate convStore
-    convStore->>convStore: Find message in activeMessages
-    convStore->>convStore: Get parent message
-    convStore->>convStore: Find sibling in parent.children[]
-    convStore->>convStore: findLeafNode(siblingId, allMessages)
-    Note right of convStore: Navigate to leaf of sibling branch
-    convStore->>convStore: updateCurrentNode(leafId)
-    convStore->>DbSvc: updateCurrentNode(convId, leafId)
-    DbSvc->>IDB: UPDATE conversations SET currNode = ?
-    convStore->>convStore: refreshActiveMessages()
-    deactivate convStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,IDB: 📝 UPDATE CONVERSATION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>convStore: updateConversationName(convId, newName)
-    activate convStore
-    convStore->>DbSvc: updateConversation(convId, {name: newName})
-    DbSvc->>IDB: UPDATE conversations SET name = ?
-    convStore->>convStore: Update in conversations array
-    deactivate convStore
-
-    Note over convStore: Auto-title update (after first response):
-    convStore->>convStore: updateConversationTitleWithConfirmation()
-    convStore->>convStore: titleUpdateConfirmationCallback?()
-    Note right of convStore: Shows dialog if title would change
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,IDB: 🗑️ DELETE CONVERSATION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>convStore: deleteConversation(convId)
-    activate convStore
-    convStore->>DbSvc: deleteConversation(convId)
-    DbSvc->>IDB: DELETE FROM conversations WHERE id = ?
-    DbSvc->>IDB: DELETE FROM messages WHERE convId = ?
-    convStore->>convStore: conversations.filter(c => c.id !== convId)
-    alt deleted active conversation
-        convStore->>convStore: clearActiveConversation()
-    end
-    deactivate convStore
-
-    UI->>convStore: deleteAll()
-    activate convStore
-    convStore->>DbSvc: Delete all conversations and messages
-    convStore->>convStore: conversations = []
-    convStore->>convStore: clearActiveConversation()
-    deactivate convStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,IDB: � MCP SERVER PER-CHAT OVERRIDES
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over convStore: Conversations can override which MCP servers are enabled.
-    Note over convStore: Uses pendingMcpServerOverrides before conversation<br/>is created, then persists to conversation metadata.
-
-    UI->>convStore: setMcpServerOverride(convId, serverName, override)
-    Note right of convStore: override = {enabled: boolean}
-
-    UI->>convStore: toggleMcpServerForChat(convId, serverName, enabled)
-    activate convStore
-    convStore->>convStore: setMcpServerOverride(convId, serverName, {enabled})
-    deactivate convStore
-
-    UI->>convStore: isMcpServerEnabledForChat(convId, serverName)
-    Note right of convStore: Check override → fall back to global MCP config
-
-    UI->>convStore: getAllMcpServerOverrides(convId)
-    Note right of convStore: Returns all overrides for a conversation
-
-    UI->>convStore: removeMcpServerOverride(convId, serverName)
-    UI->>convStore: getMcpServerOverride(convId, serverName)
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,IDB: 📤 EXPORT / 📥 IMPORT
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>convStore: exportAllConversations()
-    activate convStore
-    convStore->>DbSvc: getAllConversations()
-    loop each conversation
-        convStore->>DbSvc: getConversationMessages(convId)
-    end
-    convStore->>convStore: triggerDownload(JSON blob)
-    deactivate convStore
-
-    UI->>convStore: importConversations(file)
-    activate convStore
-    convStore->>convStore: Parse JSON file
-    convStore->>convStore: importConversationsData(parsed)
-    convStore->>DbSvc: importConversations(parsed)
-    Note right of DbSvc: Skips duplicate conversations<br/>(checks existing by ID)
-    DbSvc->>IDB: INSERT conversations + messages (skip existing)
-    convStore->>convStore: loadConversations()
-    deactivate convStore
-```
@@ -1,45 +0,0 @@
-```mermaid
-%% MODEL Mode Data Flow (single model)
-%% Detailed flows: ./flows/server-flow.mmd, ./flows/models-flow.mmd, ./flows/chat-flow.mmd
-
-sequenceDiagram
-    participant User as 👤 User
-    participant UI as 🧩 UI
-    participant Stores as 🗄️ Stores
-    participant DB as 💾 IndexedDB
-    participant API as 🌐 llama-server
-
-    Note over User,API: 🚀 Initialization (see: server-flow.mmd, models-flow.mmd)
-
-    UI->>Stores: initialize()
-    Stores->>DB: load conversations
-    Stores->>API: GET /props
-    API-->>Stores: server config + modalities
-    Stores->>API: GET /v1/models
-    API-->>Stores: single model (auto-selected)
-
-    Note over User,API: 💬 Chat Flow (see: chat-flow.mmd)
-
-    User->>UI: send message
-    UI->>Stores: sendMessage()
-    Stores->>DB: save user message
-    Stores->>API: POST /v1/chat/completions (stream)
-    loop streaming
-        API-->>Stores: SSE chunks
-        Stores-->>UI: reactive update
-    end
-    API-->>Stores: done + timings
-    Stores->>DB: save assistant message
-
-    Note over User,API: 🔁 Regenerate
-
-    User->>UI: regenerate
-    Stores->>DB: create message branch
-    Note right of Stores: same streaming flow
-
-    Note over User,API: ⏹️ Stop
-
-    User->>UI: stop
-    Stores->>Stores: abort stream
-    Stores->>DB: save partial response
-```
@@ -1,77 +0,0 @@
-```mermaid
-%% ROUTER Mode Data Flow (multi-model)
-%% Detailed flows: ./flows/server-flow.mmd, ./flows/models-flow.mmd, ./flows/chat-flow.mmd
-
-sequenceDiagram
-    participant User as 👤 User
-    participant UI as 🧩 UI
-    participant Stores as 🗄️ Stores
-    participant DB as 💾 IndexedDB
-    participant API as 🌐 llama-server
-
-    Note over User,API: 🚀 Initialization (see: server-flow.mmd, models-flow.mmd)
-
-    UI->>Stores: initialize()
-    Stores->>DB: load conversations
-    Stores->>API: GET /props
-    API-->>Stores: {role: "router"}
-    Stores->>API: GET /v1/models
-    API-->>Stores: models[] with status (loaded/available)
-    loop each loaded model
-        Stores->>API: GET /props?model=X
-        API-->>Stores: modalities (vision/audio)
-    end
-
-    Note over User,API: 🔄 Model Selection (see: models-flow.mmd)
-
-    User->>UI: select model
-    alt model not loaded
-        Stores->>API: POST /models/load
-        loop poll status
-            Stores->>API: GET /v1/models
-            API-->>Stores: check if loaded
-        end
-        Stores->>API: GET /props?model=X
-        API-->>Stores: cache modalities
-    end
-    Stores->>Stores: validate modalities vs conversation
-    alt valid
-        Stores->>Stores: select model
-    else invalid
-        Stores->>API: POST /models/unload
-        UI->>User: show error toast
-    end
-
-    Note over User,API: 💬 Chat Flow (see: chat-flow.mmd)
-
-    User->>UI: send message
-    UI->>Stores: sendMessage()
-    Stores->>DB: save user message
-    Stores->>API: POST /v1/chat/completions {model: X}
-    Note right of API: router forwards to model
-    loop streaming
-        API-->>Stores: SSE chunks + model info
-        Stores-->>UI: reactive update
-    end
-    API-->>Stores: done + timings
-    Stores->>DB: save assistant message + model used
-
-    Note over User,API: 🔁 Regenerate (optional: different model)
-
-    User->>UI: regenerate
-    Stores->>Stores: validate modalities up to this message
-    Stores->>DB: create message branch
-    Note right of Stores: same streaming flow
-
-    Note over User,API: ⏹️ Stop
-
-    User->>UI: stop
-    Stores->>Stores: abort stream
-    Stores->>DB: save partial response
-
-    Note over User,API: 🗑️ LRU Unloading
-
-    Note right of API: Server auto-unloads LRU models<br/>when cache full
-    User->>UI: select unloaded model
-    Note right of Stores: triggers load flow again
-```
@@ -1,174 +0,0 @@
-```mermaid
-sequenceDiagram
-    participant Store as 🗄️ Stores
-    participant DbSvc as ⚙️ DatabaseService
-    participant Dexie as 📦 Dexie ORM
-    participant IDB as 💾 IndexedDB
-
-    Note over DbSvc: Stateless service - all methods static<br/>Database: "LlamacppWebui"
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over Store,IDB: 📊 SCHEMA
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    rect rgb(240, 248, 255)
-        Note over IDB: conversations table:<br/>id (PK), lastModified, currNode, name
-    end
-
-    rect rgb(255, 248, 240)
-        Note over IDB: messages table:<br/>id (PK), convId (FK), type, role, timestamp,<br/>parent, children[], content, thinking,<br/>toolCalls, extra[], model, timings
-    end
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over Store,IDB: 💬 CONVERSATIONS CRUD
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Store->>DbSvc: createConversation(name)
-    activate DbSvc
-    DbSvc->>DbSvc: Generate UUID
-    DbSvc->>Dexie: db.conversations.add({id, name, lastModified, currNode: ""})
-    Dexie->>IDB: INSERT
-    IDB-->>Dexie: success
-    DbSvc-->>Store: DatabaseConversation
-    deactivate DbSvc
-
-    Store->>DbSvc: getConversation(convId)
-    DbSvc->>Dexie: db.conversations.get(convId)
-    Dexie->>IDB: SELECT WHERE id = ?
-    IDB-->>DbSvc: DatabaseConversation
-
-    Store->>DbSvc: getAllConversations()
-    DbSvc->>Dexie: db.conversations.orderBy('lastModified').reverse().toArray()
-    Dexie->>IDB: SELECT ORDER BY lastModified DESC
-    IDB-->>DbSvc: DatabaseConversation[]
-
-    Store->>DbSvc: updateConversation(convId, updates)
-    DbSvc->>Dexie: db.conversations.update(convId, {...updates, lastModified})
-    Dexie->>IDB: UPDATE
-
-    Store->>DbSvc: deleteConversation(convId)
-    activate DbSvc
-    DbSvc->>Dexie: db.conversations.delete(convId)
-    Dexie->>IDB: DELETE FROM conversations
-    DbSvc->>Dexie: db.messages.where('convId').equals(convId).delete()
-    Dexie->>IDB: DELETE FROM messages WHERE convId = ?
-    deactivate DbSvc
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over Store,IDB: 📝 MESSAGES CRUD
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Store->>DbSvc: createRootMessage(convId)
-    activate DbSvc
-    DbSvc->>DbSvc: Create root message {type: "root", parent: null}
-    DbSvc->>Dexie: db.messages.add(rootMsg)
-    Dexie->>IDB: INSERT
-    DbSvc-->>Store: rootMessageId
-    deactivate DbSvc
-
-    Store->>DbSvc: createSystemMessage(convId, content, parentId)
-    activate DbSvc
-    DbSvc->>DbSvc: Create message {role: "system", parent: parentId}
-    DbSvc->>Dexie: db.messages.add(systemMsg)
-    Dexie->>IDB: INSERT
-    DbSvc-->>Store: DatabaseMessage
-    deactivate DbSvc
-
-    Store->>DbSvc: createMessageBranch(message, parentId)
-    activate DbSvc
-    DbSvc->>DbSvc: Generate UUID for new message
-    DbSvc->>Dexie: db.messages.add({...message, id, parent: parentId})
-    Dexie->>IDB: INSERT message
-
-    alt parentId exists
-        DbSvc->>Dexie: db.messages.get(parentId)
-        Dexie->>IDB: SELECT parent
-        DbSvc->>DbSvc: parent.children.push(newId)
-        DbSvc->>Dexie: db.messages.update(parentId, {children})
-        Dexie->>IDB: UPDATE parent.children
-    end
-
-    DbSvc->>Dexie: db.conversations.update(convId, {currNode: newId})
-    Dexie->>IDB: UPDATE conversation.currNode
-    DbSvc-->>Store: DatabaseMessage
-    deactivate DbSvc
-
-    Store->>DbSvc: getConversationMessages(convId)
-    DbSvc->>Dexie: db.messages.where('convId').equals(convId).toArray()
-    Dexie->>IDB: SELECT WHERE convId = ?
-    IDB-->>DbSvc: DatabaseMessage[]
-
-    Store->>DbSvc: updateMessage(msgId, updates)
-    DbSvc->>Dexie: db.messages.update(msgId, updates)
-    Dexie->>IDB: UPDATE
-
-    Store->>DbSvc: deleteMessage(msgId)
-    DbSvc->>Dexie: db.messages.delete(msgId)
-    Dexie->>IDB: DELETE
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over Store,IDB: 🌳 BRANCHING OPERATIONS
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Store->>DbSvc: updateCurrentNode(convId, nodeId)
-    DbSvc->>Dexie: db.conversations.update(convId, {currNode: nodeId, lastModified})
-    Dexie->>IDB: UPDATE
-
-    Store->>DbSvc: deleteMessageCascading(msgId)
-    activate DbSvc
-    DbSvc->>DbSvc: findDescendantMessages(msgId, allMessages)
-    Note right of DbSvc: Recursively find all children
-    loop each descendant
-        DbSvc->>Dexie: db.messages.delete(descendantId)
-        Dexie->>IDB: DELETE
-    end
-    DbSvc->>Dexie: db.messages.delete(msgId)
-    Dexie->>IDB: DELETE target message
-
-    alt target message has a parent
-        DbSvc->>Dexie: db.messages.get(parentId)
-        DbSvc->>DbSvc: parent.children.filter(id !== msgId)
-        DbSvc->>Dexie: db.messages.update(parentId, {children})
-        Note right of DbSvc: Remove deleted message from parent's children[]
-    end
-    deactivate DbSvc
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over Store,IDB: 📥 IMPORT
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Store->>DbSvc: importConversations(data)
-    activate DbSvc
-    loop each conversation in data
-        DbSvc->>Dexie: db.conversations.get(conv.id)
-        alt conversation already exists
-            Note right of DbSvc: Skip duplicate (keep existing)
-        else conversation is new
-            DbSvc->>Dexie: db.conversations.add(conversation)
-            Dexie->>IDB: INSERT conversation
-            loop each message
-                DbSvc->>Dexie: db.messages.add(message)
-                Dexie->>IDB: INSERT message
-            end
-        end
-    end
-    deactivate DbSvc
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over Store,IDB: 🔗 MESSAGE TREE UTILITIES
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over DbSvc: Used by stores (imported from utils):
-
-    rect rgb(240, 255, 240)
-        Note over DbSvc: filterByLeafNodeId(messages, leafId)<br/>→ Returns path from root to leaf<br/>→ Used to display current branch
-    end
-
-    rect rgb(240, 255, 240)
-        Note over DbSvc: findLeafNode(startId, messages)<br/>→ Traverse to deepest child<br/>→ Used for branch navigation
-    end
-
-    rect rgb(240, 255, 240)
-        Note over DbSvc: findDescendantMessages(msgId, messages)<br/>→ Find all children recursively<br/>→ Used for cascading deletes
-    end
-```
@@ -1,226 +0,0 @@
-```mermaid
-sequenceDiagram
-    participant UI as 🧩 McpServersSettings / ChatForm
-    participant chatStore as 🗄️ chatStore
-    participant mcpStore as 🗄️ mcpStore
-    participant mcpResStore as 🗄️ mcpResourceStore
-    participant convStore as 🗄️ conversationsStore
-    participant MCPSvc as ⚙️ MCPService
-    participant LS as 💾 LocalStorage
-    participant ExtMCP as 🔌 External MCP Server
-
-    Note over mcpStore: State:<br/>isInitializing, error<br/>toolCount, connectedServers<br/>healthChecks (Map)<br/>connections (Map)<br/>toolsIndex (Map)<br/>serverConfigs (Map)
-
-    Note over mcpResStore: State:<br/>serverResources (Map)<br/>cachedResources (Map)<br/>subscriptions (Map)<br/>attachments[]
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,ExtMCP: 🚀 INITIALIZATION (App Startup)
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>mcpStore: ensureInitialized()
-    activate mcpStore
-
-    mcpStore->>LS: get(MCP_SERVERS_LOCALSTORAGE_KEY)
-    LS-->>mcpStore: MCPServerSettingsEntry[]
-
-    mcpStore->>mcpStore: parseServerSettings(servers)
-    Note right of mcpStore: Filter enabled servers<br/>Build MCPServerConfig objects<br/>Per-chat overrides checked via convStore
-
-    loop For each enabled server
-        mcpStore->>mcpStore: runHealthCheck(serverId)
-        mcpStore->>mcpStore: updateHealthCheck(id, CONNECTING)
-
-        mcpStore->>MCPSvc: connect(serverName, config, clientInfo, capabilities, onPhase)
-        activate MCPSvc
-
-        MCPSvc->>MCPSvc: createTransport(config)
-        Note right of MCPSvc: WebSocket / StreamableHTTP / SSE<br/>with optional CORS proxy
-
-        MCPSvc->>ExtMCP: Transport handshake
-        ExtMCP-->>MCPSvc: Connection established
-
-        MCPSvc->>ExtMCP: Initialize request
-        Note right of ExtMCP: Exchange capabilities<br/>Server info, protocol version
-
-        ExtMCP-->>MCPSvc: InitializeResult (serverInfo, capabilities)
-
-        MCPSvc->>ExtMCP: listTools()
-        ExtMCP-->>MCPSvc: Tool[]
-
-        MCPSvc-->>mcpStore: MCPConnection
-        deactivate MCPSvc
-
-        mcpStore->>mcpStore: connections.set(serverName, connection)
-        mcpStore->>mcpStore: indexTools(connection.tools, serverName)
-        Note right of mcpStore: toolsIndex.set(toolName, serverName)<br/>Handle name conflicts with prefixes
-
-        mcpStore->>mcpStore: updateHealthCheck(id, SUCCESS)
-        mcpStore->>mcpStore: _connectedServers.push(serverName)
-
-        alt Server supports resources
-            mcpStore->>MCPSvc: listAllResources(connection)
-            MCPSvc->>ExtMCP: listResources()
-            ExtMCP-->>MCPSvc: MCPResource[]
-            MCPSvc-->>mcpStore: resources
-
-            mcpStore->>MCPSvc: listAllResourceTemplates(connection)
-            MCPSvc->>ExtMCP: listResourceTemplates()
-            ExtMCP-->>MCPSvc: MCPResourceTemplate[]
-            MCPSvc-->>mcpStore: templates
-
-            mcpStore->>mcpResStore: setServerResources(serverName, resources, templates)
-        end
-    end
-
-    mcpStore->>mcpStore: _isInitializing = false
-    deactivate mcpStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,ExtMCP: 🔧 TOOL EXECUTION (Chat with Tools)
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>mcpStore: executeTool(mcpCall: MCPToolCall, signal?)
-    activate mcpStore
-
-    mcpStore->>mcpStore: toolsIndex.get(mcpCall.function.name)
-    Note right of mcpStore: Resolve serverName from toolsIndex<br/>MCPToolCall = {id, type, function: {name, arguments}}
-
-    mcpStore->>mcpStore: acquireConnection()
-    Note right of mcpStore: activeFlowCount++<br/>Prevent shutdown during execution
-
-    mcpStore->>mcpStore: connection = connections.get(serverName)
-
-    mcpStore->>MCPSvc: callTool(connection, {name, arguments}, signal)
-    activate MCPSvc
-
-    MCPSvc->>MCPSvc: throwIfAborted(signal)
-    MCPSvc->>ExtMCP: callTool(name, arguments)
-
-    alt Tool execution success
-        ExtMCP-->>MCPSvc: ToolCallResult (content, isError)
-        MCPSvc->>MCPSvc: formatToolResult(result)
-        Note right of MCPSvc: Handle text, image (base64),<br/>embedded resource content
-        MCPSvc-->>mcpStore: ToolExecutionResult
-    else Tool execution error
-        ExtMCP-->>MCPSvc: Error
-        MCPSvc-->>mcpStore: throw Error
-    else Aborted
-        MCPSvc-->>mcpStore: throw AbortError
-    end
-
-    deactivate MCPSvc
-
-    mcpStore->>mcpStore: releaseConnection()
-    Note right of mcpStore: activeFlowCount--
-
-    mcpStore-->>UI: ToolExecutionResult
-    deactivate mcpStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,ExtMCP: � RESOURCE ATTACHMENT CONSUMPTION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    chatStore->>mcpStore: consumeResourceAttachmentsAsExtras()
-    activate mcpStore
-    mcpStore->>mcpResStore: getAttachments()
-    mcpResStore-->>mcpStore: MCPResourceAttachment[]
-    mcpStore->>mcpStore: Convert attachments to message extras
-    mcpStore->>mcpResStore: clearAttachments()
-    mcpStore-->>chatStore: MessageExtra[] (for user message)
-    deactivate mcpStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,ExtMCP: �📝 PROMPT OPERATIONS
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>mcpStore: getAllPrompts()
-    activate mcpStore
-
-    loop For each connected server with prompts capability
-        mcpStore->>MCPSvc: listPrompts(connection)
-        MCPSvc->>ExtMCP: listPrompts()
-        ExtMCP-->>MCPSvc: Prompt[]
-        MCPSvc-->>mcpStore: prompts
-    end
-
-    mcpStore-->>UI: MCPPromptInfo[] (with serverName)
-    deactivate mcpStore
-
-    UI->>mcpStore: getPrompt(serverName, promptName, args?)
-    activate mcpStore
-
-    mcpStore->>MCPSvc: getPrompt(connection, name, args)
-    MCPSvc->>ExtMCP: getPrompt({name, arguments})
-    ExtMCP-->>MCPSvc: GetPromptResult (messages)
-    MCPSvc-->>mcpStore: GetPromptResult
-
-    mcpStore-->>UI: GetPromptResult
-    deactivate mcpStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,ExtMCP: 📁 RESOURCE OPERATIONS
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>mcpResStore: addAttachment(resourceInfo)
-    activate mcpResStore
-    mcpResStore->>mcpResStore: Create MCPResourceAttachment (loading: true)
-    mcpResStore-->>UI: attachment
-
-    UI->>mcpStore: readResource(serverName, uri)
-    activate mcpStore
-
-    mcpStore->>MCPSvc: readResource(connection, uri)
-    MCPSvc->>ExtMCP: readResource({uri})
-    ExtMCP-->>MCPSvc: MCPReadResourceResult (contents)
-    MCPSvc-->>mcpStore: contents
-
-    mcpStore-->>UI: MCPResourceContent[]
-    deactivate mcpStore
-
-    UI->>mcpResStore: updateAttachmentContent(attachmentId, content)
-    mcpResStore->>mcpResStore: cacheResourceContent(resource, content)
-    deactivate mcpResStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,ExtMCP: 🔄 AUTO-RECONNECTION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over mcpStore: On WebSocket close or connection error:
-    mcpStore->>mcpStore: autoReconnect(serverName, attempt)
-    activate mcpStore
-
-    mcpStore->>mcpStore: Calculate backoff delay
-    Note right of mcpStore: delay = min(30s, 1s * 2^attempt)
-
-    mcpStore->>mcpStore: Wait for delay
-    mcpStore->>mcpStore: reconnectServer(serverName)
-
-    alt Reconnection success
-        mcpStore->>mcpStore: updateHealthCheck(id, SUCCESS)
-    else Max attempts reached
-        mcpStore->>mcpStore: updateHealthCheck(id, ERROR)
-    end
-    deactivate mcpStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,ExtMCP: 🛑 SHUTDOWN
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>mcpStore: shutdown()
-    activate mcpStore
-
-    mcpStore->>mcpStore: Wait for activeFlowCount == 0
-
-    loop For each connection
-        mcpStore->>MCPSvc: disconnect(connection)
-        MCPSvc->>MCPSvc: transport.onclose = undefined
-        MCPSvc->>ExtMCP: close()
-    end
-
-    mcpStore->>mcpStore: connections.clear()
-    mcpStore->>mcpStore: toolsIndex.clear()
-    mcpStore->>mcpStore: _connectedServers = []
-
-    mcpStore->>mcpResStore: clear()
-    deactivate mcpStore
-```
@@ -1,181 +0,0 @@
-```mermaid
-sequenceDiagram
-    participant UI as 🧩 ModelsSelector
-    participant Hooks as 🪝 useModelChangeValidation
-    participant modelsStore as 🗄️ modelsStore
-    participant serverStore as 🗄️ serverStore
-    participant convStore as 🗄️ conversationsStore
-    participant ModelsSvc as ⚙️ ModelsService
-    participant PropsSvc as ⚙️ PropsService
-    participant API as 🌐 llama-server
-
-    Note over modelsStore: State:<br/>models: ModelOption[]<br/>routerModels: ApiModelDataEntry[]<br/>selectedModelId, selectedModelName<br/>loading, updating, error<br/>modelLoadingStates (Map)<br/>modelPropsCache (Map)<br/>propsCacheVersion
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 🚀 INITIALIZATION (MODEL mode)
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>modelsStore: fetch()
-    activate modelsStore
-    modelsStore->>modelsStore: loading = true
-
-    alt serverStore.props not loaded
-        modelsStore->>serverStore: fetch()
-        Note over serverStore: → see server-flow.mmd
-    end
-
-    modelsStore->>ModelsSvc: list()
-    ModelsSvc->>API: GET /v1/models
-    API-->>ModelsSvc: ApiModelListResponse {data: [model]}
-
-    modelsStore->>modelsStore: models = $state(mapped)
-    Note right of modelsStore: Map to ModelOption[]:<br/>{id, name, model, description, capabilities}
-
-    Note over modelsStore: MODEL mode: Get modalities from serverStore.props
-    modelsStore->>modelsStore: modelPropsCache.set(model.id, serverStore.props)
-    modelsStore->>modelsStore: models[0].modalities = props.modalities
-
-    modelsStore->>modelsStore: Auto-select single model
-    Note right of modelsStore: selectedModelId = models[0].id
-    modelsStore->>modelsStore: loading = false
-    deactivate modelsStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 🚀 INITIALIZATION (ROUTER mode)
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>modelsStore: fetch()
-    activate modelsStore
-    modelsStore->>ModelsSvc: list()
-    ModelsSvc->>API: GET /v1/models
-    API-->>ModelsSvc: ApiModelListResponse
-    modelsStore->>modelsStore: models = $state(mapped)
-    deactivate modelsStore
-
-    Note over UI: After models loaded, layout triggers:
-    UI->>modelsStore: fetchRouterModels()
-    activate modelsStore
-    modelsStore->>ModelsSvc: listRouter()
-    ModelsSvc->>API: GET /v1/models
-    API-->>ModelsSvc: ApiRouterModelsListResponse
-    Note right of API: {data: [{id, status, path, in_cache}]}
-    modelsStore->>modelsStore: routerModels = $state(data)
-
-    modelsStore->>modelsStore: fetchModalitiesForLoadedModels()
-    loop each model where status === "loaded"
-        modelsStore->>PropsSvc: fetchForModel(modelId)
-        PropsSvc->>API: GET /props?model={modelId}
-        API-->>PropsSvc: ApiLlamaCppServerProps
-        modelsStore->>modelsStore: modelPropsCache.set(modelId, props)
-    end
-    modelsStore->>modelsStore: propsCacheVersion++
-    deactivate modelsStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 🔄 MODEL SELECTION (ROUTER mode)
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>Hooks: useModelChangeValidation({getRequiredModalities, onSuccess?, onValidationFailure?})
-    Note over Hooks: Hook configured per-component:<br/>ChatForm: getRequiredModalities = usedModalities<br/>ChatMessage: getRequiredModalities = getModalitiesUpToMessage(msgId)
-
-    UI->>Hooks: handleModelChange(modelId, modelName)
-    activate Hooks
-    Hooks->>Hooks: previousSelectedModelId = modelsStore.selectedModelId
-    Hooks->>modelsStore: isModelLoaded(modelName)?
-
-    alt model NOT loaded
-        Hooks->>modelsStore: loadModel(modelName)
-        Note over modelsStore: → see LOAD MODEL section below
-    end
-
-    Note over Hooks: Always fetch props (from cache or API)
-    Hooks->>modelsStore: fetchModelProps(modelName)
-    modelsStore-->>Hooks: props
-
-    Hooks->>convStore: getRequiredModalities()
-    convStore-->>Hooks: {vision, audio}
-
-    Hooks->>Hooks: Validate: model.modalities ⊇ required?
-
-    alt validation PASSED
-        Hooks->>modelsStore: selectModelById(modelId)
-        Hooks-->>UI: return true
-    else validation FAILED
-        Hooks->>UI: toast.error("Model doesn't support required modalities")
-        alt model was just loaded
-            Hooks->>modelsStore: unloadModel(modelName)
-        end
-        alt onValidationFailure provided
-            Hooks->>modelsStore: selectModelById(previousSelectedModelId)
-        end
-        Hooks-->>UI: return false
-    end
-    deactivate Hooks
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: ⬆️ LOAD MODEL (ROUTER mode)
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    modelsStore->>modelsStore: loadModel(modelId)
-    activate modelsStore
-
-    alt already loaded
-        modelsStore-->>modelsStore: return (no-op)
-    end
-
-    modelsStore->>modelsStore: modelLoadingStates.set(modelId, true)
-    modelsStore->>ModelsSvc: load(modelId)
-    ModelsSvc->>API: POST /models/load {model: modelId}
-    API-->>ModelsSvc: {status: "loading"}
-
-    modelsStore->>modelsStore: pollForModelStatus(modelId, LOADED)
-    loop poll every 500ms (max 60 attempts)
-        modelsStore->>modelsStore: fetchRouterModels()
-        modelsStore->>ModelsSvc: listRouter()
-        ModelsSvc->>API: GET /v1/models
-        API-->>ModelsSvc: models[]
-        modelsStore->>modelsStore: getModelStatus(modelId)
-        alt status === LOADED
-            Note right of modelsStore: break loop
-        else status === LOADING
-            Note right of modelsStore: wait 500ms, continue
-        end
-    end
-
-    modelsStore->>modelsStore: updateModelModalities(modelId)
-    modelsStore->>PropsSvc: fetchForModel(modelId)
-    PropsSvc->>API: GET /props?model={modelId}
-    API-->>PropsSvc: props with modalities
-    modelsStore->>modelsStore: modelPropsCache.set(modelId, props)
-    modelsStore->>modelsStore: propsCacheVersion++
-
-    modelsStore->>modelsStore: modelLoadingStates.set(modelId, false)
-    deactivate modelsStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: ⬇️ UNLOAD MODEL (ROUTER mode)
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    modelsStore->>modelsStore: unloadModel(modelId)
-    activate modelsStore
-    modelsStore->>modelsStore: modelLoadingStates.set(modelId, true)
-    modelsStore->>ModelsSvc: unload(modelId)
-    ModelsSvc->>API: POST /models/unload {model: modelId}
-
-    modelsStore->>modelsStore: pollForModelStatus(modelId, UNLOADED)
-    loop poll until unloaded
-        modelsStore->>ModelsSvc: listRouter()
-        ModelsSvc->>API: GET /v1/models
-    end
-
-    modelsStore->>modelsStore: modelLoadingStates.set(modelId, false)
-    deactivate modelsStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 📊 COMPUTED GETTERS
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over modelsStore: Getters:<br/>- selectedModel: ModelOption | null<br/>- loadedModelIds: string[] (from routerModels)<br/>- loadingModelIds: string[] (from modelLoadingStates)<br/>- singleModelName: string | null (MODEL mode only)
-
-    Note over modelsStore: Modality helpers:<br/>- getModelModalities(modelId): {vision, audio}<br/>- modelSupportsVision(modelId): boolean<br/>- modelSupportsAudio(modelId): boolean
-```
@@ -1,76 +0,0 @@
-```mermaid
-sequenceDiagram
-    participant UI as 🧩 +layout.svelte
-    participant serverStore as 🗄️ serverStore
-    participant PropsSvc as ⚙️ PropsService
-    participant API as 🌐 llama-server
-
-    Note over serverStore: State:<br/>props: ApiLlamaCppServerProps | null<br/>loading, error<br/>role: ServerRole | null (MODEL | ROUTER)<br/>fetchPromise (deduplication)
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 🚀 INITIALIZATION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>serverStore: fetch()
-    activate serverStore
-
-    alt fetchPromise exists (already fetching)
-        serverStore-->>UI: return fetchPromise
-        Note right of serverStore: Deduplicate concurrent calls
-    end
-
-    serverStore->>serverStore: loading = true
-    serverStore->>serverStore: fetchPromise = new Promise()
-
-    serverStore->>PropsSvc: fetch()
-    PropsSvc->>API: GET /props
-    API-->>PropsSvc: ApiLlamaCppServerProps
-    Note right of API: {role, model_path, model_alias,<br/>modalities, default_generation_settings, ...}
-
-    PropsSvc-->>serverStore: props
-    serverStore->>serverStore: props = $state(data)
-
-    serverStore->>serverStore: detectRole(props)
-    Note right of serverStore: role = props.role === "router"<br/>  ? ServerRole.ROUTER<br/>  : ServerRole.MODEL
-
-    serverStore->>serverStore: loading = false
-    serverStore->>serverStore: fetchPromise = null
-    deactivate serverStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 📊 COMPUTED GETTERS
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over serverStore: Getters from props:
-
-    rect rgb(240, 255, 240)
-        Note over serverStore: defaultParams<br/>→ props.default_generation_settings.params<br/>(temperature, top_p, top_k, etc.)
-    end
-
-    rect rgb(240, 255, 240)
-        Note over serverStore: contextSize<br/>→ props.default_generation_settings.n_ctx
-    end
-
-    rect rgb(255, 240, 240)
-        Note over serverStore: isRouterMode<br/>→ role === ServerRole.ROUTER
-    end
-
-    rect rgb(255, 240, 240)
-        Note over serverStore: isModelMode<br/>→ role === ServerRole.MODEL
-    end
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: 🔗 RELATIONSHIPS
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over serverStore: Used by:
-    Note right of serverStore: - modelsStore: role detection, MODEL mode modalities<br/>- settingsStore: syncWithServerDefaults (defaultParams)<br/>- chatStore: contextSize for processing state<br/>- UI components: isRouterMode for conditional rendering
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,API: ❌ ERROR HANDLING
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over serverStore: getErrorMessage(): string | null<br/>Returns formatted error for UI display
-
-    Note over serverStore: clear(): void<br/>Resets all state (props, error, loading, role)
-```
@@ -1,156 +0,0 @@
-```mermaid
-sequenceDiagram
-    participant UI as 🧩 ChatSettings
-    participant settingsStore as 🗄️ settingsStore
-    participant serverStore as 🗄️ serverStore
-    participant ParamSvc as ⚙️ ParameterSyncService
-    participant LS as 💾 LocalStorage
-
-    Note over settingsStore: State:<br/>config: SettingsConfigType<br/>theme: string ("auto" | "light" | "dark")<br/>isInitialized: boolean<br/>userOverrides: Set&lt;string&gt;
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,LS: 🚀 INITIALIZATION
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over settingsStore: Auto-initialized in constructor (browser only)
-    settingsStore->>settingsStore: initialize()
-    activate settingsStore
-
-    settingsStore->>settingsStore: loadConfig()
-    settingsStore->>LS: get("llama-config")
-    LS-->>settingsStore: StoredConfig | null
-
-    alt config exists
-        settingsStore->>settingsStore: Merge with SETTING_CONFIG_DEFAULT
-        Note right of settingsStore: Fill missing keys with defaults
-    else no config
-        settingsStore->>settingsStore: config = SETTING_CONFIG_DEFAULT
-    end
-
-    settingsStore->>LS: get("llama-userOverrides")
-    LS-->>settingsStore: string[] | null
-    settingsStore->>settingsStore: userOverrides = new Set(data)
-
-    settingsStore->>settingsStore: loadTheme()
-    settingsStore->>LS: get("llama-theme")
-    LS-->>settingsStore: theme | "auto"
-
-    settingsStore->>settingsStore: isInitialized = true
-    deactivate settingsStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,LS: 🔄 SYNC WITH SERVER DEFAULTS
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over UI: Triggered from +layout.svelte when serverStore.props loaded
-    UI->>settingsStore: syncWithServerDefaults()
-    activate settingsStore
-
-    settingsStore->>serverStore: defaultParams
-    serverStore-->>settingsStore: {temperature, top_p, top_k, ...}
-
-    loop each SYNCABLE_PARAMETER
-        alt key NOT in userOverrides
-            settingsStore->>settingsStore: config[key] = serverDefault[key]
-            Note right of settingsStore: Non-overridden params adopt server default
-        else key in userOverrides
-            Note right of settingsStore: Keep user value, skip server default
-        end
-    end
-
-    alt serverStore.props has uiSettings
-        settingsStore->>settingsStore: Apply uiSettings from server
-        Note right of settingsStore: Server-provided UI settings<br/>(e.g. showRawOutputSwitch)
-    end
-
-    settingsStore->>settingsStore: saveConfig()
-    deactivate settingsStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,LS: ⚙️ UPDATE CONFIG
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>settingsStore: updateConfig(key, value)
-    activate settingsStore
-    settingsStore->>settingsStore: config[key] = value
-
-    alt value matches server default for key
-        settingsStore->>settingsStore: userOverrides.delete(key)
-        Note right of settingsStore: Matches server default, remove override
-    else value differs from server default
-        settingsStore->>settingsStore: userOverrides.add(key)
-        Note right of settingsStore: Mark as user-modified (won't be overwritten)
-    end
-
-    settingsStore->>settingsStore: saveConfig()
-    settingsStore->>LS: set(CONFIG_LOCALSTORAGE_KEY, config)
-    settingsStore->>LS: set(USER_OVERRIDES_LOCALSTORAGE_KEY, [...userOverrides])
-    deactivate settingsStore
-
-    UI->>settingsStore: updateMultipleConfig({key1: val1, key2: val2})
-    activate settingsStore
-    Note right of settingsStore: Batch update, single save
-    settingsStore->>settingsStore: For each key: config[key] = value
-    settingsStore->>settingsStore: For each key: userOverrides.add(key)
-    settingsStore->>settingsStore: saveConfig()
-    deactivate settingsStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,LS: 🔄 RESET
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>settingsStore: resetConfig()
-    activate settingsStore
-    settingsStore->>settingsStore: config = {...SETTING_CONFIG_DEFAULT}
-    settingsStore->>settingsStore: userOverrides.clear()
-    Note right of settingsStore: All params reset to defaults<br/>Next syncWithServerDefaults will adopt server values
-    settingsStore->>settingsStore: saveConfig()
-    deactivate settingsStore
-
-    UI->>settingsStore: resetParameterToServerDefault(key)
-    activate settingsStore
-    settingsStore->>settingsStore: userOverrides.delete(key)
-    settingsStore->>serverStore: defaultParams[key]
-    settingsStore->>settingsStore: config[key] = serverDefault
-    settingsStore->>settingsStore: saveConfig()
-    deactivate settingsStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,LS: 🎨 THEME
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>settingsStore: updateTheme(newTheme)
-    activate settingsStore
-    settingsStore->>settingsStore: theme = newTheme
-    settingsStore->>settingsStore: saveTheme()
-    settingsStore->>LS: set("llama-theme", theme)
-    deactivate settingsStore
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,LS: 📊 PARAMETER INFO
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    UI->>settingsStore: getParameterInfo(key)
-    settingsStore->>ParamSvc: getParameterInfo(key, config, serverDefaults, userOverrides)
-    ParamSvc-->>settingsStore: ParameterInfo
-    Note right of ParamSvc: {<br/>  currentValue,<br/>  serverDefault,<br/>  isUserOverride: boolean,<br/>  canSync: boolean,<br/>  isDifferentFromServer: boolean<br/>}
-
-    UI->>settingsStore: getParameterDiff()
-    settingsStore->>ParamSvc: createParameterDiff(config, serverDefaults, userOverrides)
-    ParamSvc-->>settingsStore: ParameterDiff[]
-    Note right of ParamSvc: Array of parameters where user != server
-
-    %% ═══════════════════════════════════════════════════════════════════════════
-    Note over UI,LS: 📋 CONFIG CATEGORIES
-    %% ═══════════════════════════════════════════════════════════════════════════
-
-    Note over settingsStore: Syncable with server (from /props):
-    rect rgb(240, 255, 240)
-        Note over settingsStore: temperature, top_p, top_k, min_p<br/>repeat_penalty, presence_penalty, frequency_penalty<br/>dynatemp_range, dynatemp_exponent<br/>typ_p, xtc_probability, xtc_threshold<br/>dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n
-    end
-
-    Note over settingsStore: UI-only (not synced):
-    rect rgb(255, 240, 240)
-        Note over settingsStore: systemMessage, custom (JSON)<br/>showStatistics, enableContinueGeneration<br/>autoMicOnEmpty, disableAutoScroll<br/>apiKey, pdfAsImage, disableReasoningParsing, showRawOutputSwitch
-    end
-```
@@ -12,6 +12,49 @@ import { fileURLToPath } from 'node:url';
 import ts from 'typescript-eslint';

 const gitignorePath = fileURLToPath(new URL('./.gitignore', import.meta.url));
+// Require a blank line between consecutive class accessors (get/set). The core
+// `padding-line-between-statements` rule only handles statements, not class
+// members, so this is enforced with a small custom rule.
+const blankLineBetweenAccessors = {
+	create(context) {
+		return {
+			MethodDefinition(node) {
+				if (node.kind !== 'get' && node.kind !== 'set') return;
+
+				const body = node.parent;
+
+				if (!body || body.type !== 'ClassBody') return;
+
+				const index = body.body.indexOf(node);
+
+				if (index <= 0) return;
+
+				const prev = body.body[index - 1];
+
+				if (prev.type !== 'MethodDefinition' || (prev.kind !== 'get' && prev.kind !== 'set'))
+					return;
+
+				if (node.loc.start.line - prev.loc.end.line <= 1) {
+					context.report({
+						fix(fixer) {
+							// Insert after the previous accessor's closing brace so the blank
+							// line keeps the current accessor's indentation.
+							return fixer.insertTextAfter(prev, '\n');
+						},
+						message: 'Expected a blank line between class accessors (get/set).',
+						node
+					});
+				}
+			}
+		};
+	},
+	meta: {
+		docs: { description: 'Require a blank line between consecutive class accessors (get/set).' },
+		fixable: 'whitespace',
+		schema: [],
+		type: 'layout'
+	}
+};

 export default ts.config(
 	includeIgnoreFile(gitignorePath),
@@ -22,7 +65,11 @@ export default ts.config(
 	...svelte.configs.prettier,
 	{
 		languageOptions: { globals: { ...globals.browser, ...globals.node } },
-		plugins: { perfectionist, 'simple-import-sort': simpleImportSort },
+		plugins: {
+			local: { rules: { 'blank-line-between-accessors': blankLineBetweenAccessors } },
+			perfectionist,
+			'simple-import-sort': simpleImportSort
+		},
 		rules: {
 			// Snippet bodies often ignore one or more of the parent's params
 			// (e.g. `{#snippet children(_meta, ctx)}` when only ctx is read).
@@ -30,8 +77,11 @@ export default ts.config(
 				'error',
 				{ argsIgnorePattern: '^_', varsIgnorePattern: '^_' }
 			],
+
 			// Enforce empty line at end of file
 			'eol-last': 'error',
+			// Enforce a blank line between consecutive get/set accessors
+			'local/blank-line-between-accessors': 'error',
 			// typescript-eslint strongly recommend that you do not use the no-undef lint rule on TypeScript projects.
 			// see: https://typescript-eslint.io/troubleshooting/faqs/eslint/#i-get-errors-from-the-no-undef-rule-about-global-variables-not-being-defined-even-though-there-are-no-typescript-errors
 			'no-undef': 'off',
@@ -61,6 +111,38 @@ export default ts.config(
 				{ blankLine: 'always', next: ['return', 'throw', 'break', 'continue'], prev: '*' }
 			],

+			// Class member order: public fields -> private fields -> constructor -> getters
+			// -> setters -> public methods -> private methods, alphabetical within each.
+			// Svelte $derived fields must stay in dependency order (forward references are
+			// rejected), so the two stores that rely on that are exempted below.
+			'perfectionist/sort-classes': [
+				'error',
+				{
+					customGroups: [
+						{ groupName: 'public-field', modifiers: ['public'], selector: 'property' },
+						{ groupName: 'private-field', modifiers: ['private'], selector: 'property' },
+						{ groupName: 'get-method', selector: 'get-method' },
+						{ groupName: 'set-method', selector: 'set-method' },
+						{ groupName: 'public-method', modifiers: ['public'], selector: 'method' },
+						{ groupName: 'private-method', modifiers: ['private'], selector: 'method' }
+					],
+					groups: [
+						'public-field',
+						'private-field',
+						'constructor',
+						'get-method',
+						'set-method',
+						'public-method',
+						'private-method',
+						'unknown'
+					],
+					type: 'natural',
+					// Keep members in dependency order (Svelte rejects forward references in
+					// $derived fields), while still sorting the rest alphabetically.
+					useExperimentalDependencyDetection: true
+				}
+			],
+
 			// Alphabetical order for enum members
 			'perfectionist/sort-enums': ['error', { type: 'natural' }],

@@ -139,7 +139,7 @@
 	let fileSize = $derived(currentItem?.size ? formatFileSize(currentItem.size) : '');

 	let hasVisionModality = $derived(
-		currentItem && activeModelId ? modelsStore.modelSupportsVision(activeModelId) : false
+		currentItem && activeModelId ? modelsStore.props.modelSupportsVision(activeModelId) : false
 	);

 	let audioSrc = $derived(
@@ -28,7 +28,6 @@
 	import {
 		chatStore,
 		conversationsStore,
-		mcpResourceStore,
 		mcpStore,
 		modelsStore,
 		serverStore,
@@ -140,7 +139,9 @@
 	// float above the box.
 	let mentionAnchor: HTMLDivElement | null = $state(null);

-	let cwd = $derived(conversationsStore.activeConversation?.cwd ?? conversationsStore.pendingCwd);
+	let cwd = $derived(
+		conversationsStore.activeConversation?.cwd ?? conversationsStore.preferences.pendingCwd
+	);

 	const pickers = useChatFormPickers({
 		focusInput: refocusInput,
@@ -151,7 +152,8 @@
 		getShowModelSelector: () => showModelSelector,
 		getValue: () => value,
 		hasCwdTools: () => toolsStore.hasEnabledCwdTools,
-		hasPrompts: () => mcpStore.hasPromptsCapability(conversationsStore.getAllMcpServerOverrides()),
+		hasPrompts: () =>
+			mcpStore.hasPromptsCapability(conversationsStore.preferences.getAllMcpServerOverrides()),
 		openModelSelector: () => chatFormActionsRef?.openModelSelector(),
 		setCaretOffset: (offset) => inputRef?.setCaretOffset(offset),
 		setValue: (v) => {
@@ -170,7 +172,7 @@
 			onValueChange?.('');
 		}

-		await conversationsStore.setCwd(newDir);
+		await conversationsStore.preferences.setCwd(newDir);

 		if (conversationsStore.activeConversation) {
 			await chatStore.recordCwdChange(newDir?.trim() || null);
@@ -595,7 +597,7 @@
 				{useRichInput}
 			/>

-			{#if mcpResourceStore.hasAttachments}
+			{#if mcpStore.resources.hasAttachments}
 				<ChatFormMcpResourcesList
 					class="mb-3"
 					onResourceClick={(uri) => {
@@ -38,11 +38,11 @@
 	}

 	function isServerEnabledForChat(serverId: string): boolean {
-		return conversationsStore.isMcpServerEnabledForChat(serverId);
+		return conversationsStore.preferences.isMcpServerEnabledForChat(serverId);
 	}

 	async function toggleServerForChat(serverId: string) {
-		await conversationsStore.toggleMcpServerForChat(serverId);
+		await conversationsStore.preferences.toggleMcpServerForChat(serverId);
 	}

 	function handleMcpSubMenuOpen(open: boolean) {
@@ -218,12 +218,15 @@
 								{@const hasError = healthState.status === HealthCheckStatus.ERROR}
 								{@const displayName = mcpStore.getServerLabel(server)}
 								{@const faviconUrl = mcpStore.getServerFavicon(server.id)}
-								{@const isEnabled = conversationsStore.isMcpServerEnabledForChat(server.id)}
+								{@const isEnabled = conversationsStore.preferences.isMcpServerEnabledForChat(
+									server.id
+								)}

 								<button
 									type="button"
 									class={sheetItemRowClass}
-									onclick={() => !hasError && conversationsStore.toggleMcpServerForChat(server.id)}
+									onclick={() =>
+										!hasError && conversationsStore.preferences.toggleMcpServerForChat(server.id)}
 									disabled={hasError}
 								>
 									<div class="flex min-w-0 flex-1 items-center gap-2">
@@ -250,7 +253,8 @@
 									{:else}
 										<Switch
 											checked={isEnabled}
-											onCheckedChange={() => conversationsStore.toggleMcpServerForChat(server.id)}
+											onCheckedChange={() =>
+												conversationsStore.preferences.toggleMcpServerForChat(server.id)}
 										/>
 									{/if}
 								</button>
@@ -81,10 +81,10 @@

 	$effect(() => {
 		if (activeModelId) {
-			const cached = modelsStore.getModelProps(activeModelId);
+			const cached = modelsStore.props.getModelProps(activeModelId);

 			if (!cached) {
-				modelsStore.fetchModelProps(activeModelId).then(() => {
+				modelsStore.props.fetchModelProps(activeModelId).then(() => {
 					modelPropsVersion++;
 				});
 			}
@@ -94,19 +94,21 @@
 	$effect(() => {
 		void modelPropsVersion;

-		hasAudioModality = activeModelId ? modelsStore.modelSupportsAudio(activeModelId) : false;
+		hasAudioModality = activeModelId ? modelsStore.props.modelSupportsAudio(activeModelId) : false;
 	});

 	$effect(() => {
 		void modelPropsVersion;

-		hasVideoModality = activeModelId ? modelsStore.modelSupportsVideo(activeModelId) : false;
+		hasVideoModality = activeModelId ? modelsStore.props.modelSupportsVideo(activeModelId) : false;
 	});

 	$effect(() => {
 		void modelPropsVersion;

-		hasVisionModality = activeModelId ? modelsStore.modelSupportsVision(activeModelId) : false;
+		hasVisionModality = activeModelId
+			? modelsStore.props.modelSupportsVision(activeModelId)
+			: false;
 	});

 	$effect(() => {
@@ -58,13 +58,13 @@
 	let currentConfig = $derived(settingsStore.config);

 	let hasMcpPromptsSupport = $derived.by(() => {
-		const perChatOverrides = conversationsStore.getAllMcpServerOverrides();
+		const perChatOverrides = conversationsStore.preferences.getAllMcpServerOverrides();

 		return mcpStore.hasPromptsCapability(perChatOverrides);
 	});

 	let hasMcpResourcesSupport = $derived.by(() => {
-		const perChatOverrides = conversationsStore.getAllMcpServerOverrides();
+		const perChatOverrides = conversationsStore.preferences.getAllMcpServerOverrides();

 		return mcpStore.hasResourcesCapability(perChatOverrides);
 	});
@@ -121,7 +121,7 @@

 		if (!chatStore.isLoading && !chatStore.isStreaming()) return false;

-		const processingState = chatStore.activeProcessingState;
+		const processingState = chatStore.processing.activeState;

 		if (!processingState) return false;

@@ -16,7 +16,7 @@
 	$effect(() => {
 		const conv = conversationsStore.activeConversation;

-		untrack(() => chatStore.setActiveProcessingConversation(conv?.id ?? null));
+		untrack(() => chatStore.processing.setActiveConversation(conv?.id ?? null));
 	});

 	$effect(() => {
@@ -28,12 +28,12 @@
 		if (chatStore.isLoading || chatStore.isStreaming()) return;

 		if (messages.length === 0) {
-			untrack(() => chatStore.clearProcessingState(conv.id));
+			untrack(() => chatStore.processing.setState(conv.id, null));

 			return;
 		}

-		untrack(() => chatStore.restoreProcessingStateFromMessages(messages, conv.id));
+		untrack(() => chatStore.processing.restoreFromMessages(messages, conv.id));
 	});

 	$effect(() => {
@@ -3,7 +3,7 @@
 		ChatAttachmentsListItemMcpResource,
 		HorizontalScrollCarousel
 	} from '$lib/components/app';
-	import { mcpResourceStore, mcpStore } from '$lib/stores';
+	import { mcpStore } from '$lib/stores';

 	interface Props {
 		class?: string;
@@ -12,8 +12,8 @@

 	let { class: className, onResourceClick }: Props = $props();

-	const attachments = $derived(mcpResourceStore.attachments);
-	const hasAttachments = $derived(mcpResourceStore.hasAttachments);
+	const attachments = $derived(mcpStore.resources.attachments);
+	const hasAttachments = $derived(mcpStore.resources.hasAttachments);

 	function handleRemove(attachmentId: string) {
 		mcpStore.removeResourceAttachment(attachmentId);
@@ -87,7 +87,7 @@
 		isLoading = true;

 		try {
-			const perChatOverrides = conversationsStore.getAllMcpServerOverrides();
+			const perChatOverrides = conversationsStore.preferences.getAllMcpServerOverrides();
 			const initialized = await mcpStore.ensureInitialized(perChatOverrides);

 			if (!initialized) {
@@ -59,7 +59,7 @@
 		message.model ?? chatStore.getResumeModel(message.convId) ?? modelsStore.selectedModelName
 	);
 	let modelLoadProgress = $derived(
-		isRouter && loadTargetModel ? modelsStore.getLoadProgress(loadTargetModel) : null
+		isRouter && loadTargetModel ? modelsStore.status.getLoadProgress(loadTargetModel) : null
 	);
 	let modelLoadingText = $derived(modelLoadProgressText(modelLoadProgress));

@@ -31,7 +31,7 @@
 				pendingModel = modelId;

 				try {
-					await modelsStore.loadModel(modelId);
+					await modelsStore.status.load(modelId);
 				} finally {
 					pendingModel = null;
 				}
@@ -43,14 +43,14 @@
 	);

 	const hasReasoningError = $derived(
-		isLastAssistantMessage ? !!agenticStore.lastError(message.convId) : false
+		isLastAssistantMessage ? !!agenticStore.getLastError(message.convId) : false
 	);

 	let permissionDismissed = $state(false);

 	const pendingPermission = $derived(
 		isStreaming && isLastAssistantMessage
-			? agenticStore.pendingPermissionRequest(message.convId)
+			? agenticStore.getPendingPermissionRequest(message.convId)
 			: null
 	);

@@ -74,7 +74,7 @@

 	const pendingContinue = $derived(
 		isStreaming && isLastAssistantMessage
-			? agenticStore.pendingContinueRequest(message.convId)
+			? agenticStore.getPendingContinueRequest(message.convId)
 			: false
 	);

@@ -97,7 +97,7 @@
 	const sections = $derived(deriveAgenticSections(message, toolMessages, [], isStreaming));

 	const currentlyExecutingToolCallId = $derived(
-		isStreaming ? agenticStore.executingToolCallId(message.convId) : null
+		isStreaming ? agenticStore.getExecutingToolCallId(message.convId) : null
 	);

 	type TurnGroup = {
@@ -238,30 +238,30 @@
 		/>
 	{/each}

-	{#if conversationsStore.activeConversation && agenticStore.pendingSteeringMessageContent(conversationsStore.activeConversation!.id)}
+	{#if conversationsStore.activeConversation && agenticStore.getPendingSteeringMessageContent(conversationsStore.activeConversation!.id)}
 		{@const convId = conversationsStore.activeConversation!.id}
-		{@const pendingContent = agenticStore.pendingSteeringMessageContent(convId)}
+		{@const pendingContent = agenticStore.getPendingSteeringMessageContent(convId)}

 		{#if pendingContent}
 			<ChatMessageUserPending
 				class="mx-auto mt-12 w-full max-w-[48rem]"
 				content={pendingContent}
-				extras={agenticStore.pendingSteeringMessageExtras(convId)}
+				extras={agenticStore.getPendingSteeringMessageExtras(convId)}
 				onSendImmediately={() => chatStore.abortCurrentFlow(convId)}
 				onEdit={(newContent, extras) =>
 					agenticStore.injectSteeringMessage(convId, newContent, extras)}
 				onDelete={() => agenticStore.clearSteeringMessage(convId)}
 			/>
 		{/if}
-	{:else if conversationsStore.activeConversation && chatStore.pendingMessageContent(conversationsStore.activeConversation!.id)}
+	{:else if conversationsStore.activeConversation && chatStore.getPendingMessageContent(conversationsStore.activeConversation!.id)}
 		{@const convId = conversationsStore.activeConversation!.id}
-		{@const pendingContent = chatStore.pendingMessageContent(convId)}
+		{@const pendingContent = chatStore.getPendingMessageContent(convId)}

 		{#if pendingContent}
 			<ChatMessageUserPending
 				class="mx-auto mt-12 w-full max-w-[48rem]"
 				content={pendingContent}
-				extras={chatStore.pendingMessageExtras(convId)}
+				extras={chatStore.getPendingMessageExtras(convId)}
 				onSendImmediately={() => chatStore.abortCurrentFlow(convId)}
 				onEdit={(newContent, extras) => chatStore.injectPendingMessage(convId, newContent, extras)}
 				onDelete={() => chatStore.clearPendingMessage(convId)}
@@ -8,7 +8,7 @@
 	import { Button } from '$lib/components/ui/button';
 	import * as Dialog from '$lib/components/ui/dialog';
 	import { ICON_CLASS_DEFAULT } from '$lib/constants';
-	import { conversationsStore, mcpResourceStore, mcpStore } from '$lib/stores';
+	import { conversationsStore, mcpStore } from '$lib/stores';
 	import type { MCPResourceContent, MCPResourceInfo, MCPResourceTemplateInfo } from '$lib/types';
 	import { getResourceDisplayName } from '$lib/utils';
 	import { SvelteSet } from 'svelte/reactivity';
@@ -33,7 +33,7 @@
 	let templatePreviewLoading = $state(false);
 	let templatePreviewError = $state<string | null>(null);

-	const totalCount = $derived(mcpResourceStore.totalResourceCount);
+	const totalCount = $derived(mcpStore.resources.totalResourceCount);

 	$effect(() => {
 		if (open) {
@@ -48,7 +48,7 @@
 	});

 	async function loadResources() {
-		const perChatOverrides = conversationsStore.getAllMcpServerOverrides();
+		const perChatOverrides = conversationsStore.preferences.getAllMcpServerOverrides();
 		const initialized = await mcpStore.ensureInitialized(perChatOverrides);

 		if (initialized) {
@@ -126,16 +126,16 @@
 		isAttaching = true;

 		try {
-			const knownResource = mcpResourceStore.findResourceByUri(templatePreviewUri);
+			const knownResource = mcpStore.resources.findResourceByUri(templatePreviewUri);

 			if (knownResource) {
-				if (!mcpResourceStore.isAttached(knownResource.uri)) {
+				if (!mcpStore.resources.isAttached(knownResource.uri)) {
 					await mcpStore.attachResource(knownResource.uri);
 				}

 				toast.success(`Resource attached: ${knownResource.title || knownResource.name}`);
 			} else {
-				if (mcpResourceStore.isAttached(templatePreviewUri)) {
+				if (mcpStore.resources.isAttached(templatePreviewUri)) {
 					toast.info('Resource already attached');
 					handleOpenChange(false);

@@ -147,9 +147,9 @@
 					serverName: selectedTemplate.serverName,
 					uri: templatePreviewUri
 				};
-				const attachment = mcpResourceStore.addAttachment(resourceInfo);
+				const attachment = mcpStore.resources.addAttachment(resourceInfo);

-				mcpResourceStore.updateAttachmentContent(attachment.id, templatePreviewContent);
+				mcpStore.resources.updateAttachmentContent(attachment.id, templatePreviewContent);

 				toast.success(`Resource attached: ${resourceInfo.name}`);
 			}
@@ -199,7 +199,7 @@

 	function getAllResourcesFlatInTreeOrder(): MCPResourceInfo[] {
 		const allResources: MCPResourceInfo[] = [];
-		const resourcesMap = mcpResourceStore.serverResources;
+		const resourcesMap = mcpStore.resources.serverResources;

 		for (const [serverName, serverRes] of resourcesMap.entries()) {
 			for (const resource of serverRes.resources) {
@@ -234,7 +234,7 @@
 			useProxy: newServerUseProxy
 		});

-		conversationsStore.setMcpServerOverride(newServerId, true);
+		conversationsStore.preferences.setMcpServerOverride(newServerId, true);

 		handleOpenChange(false);
 	}
@@ -42,7 +42,7 @@
 	let modalities = $derived.by(() => {
 		if (!firstModel?.id) return [];

-		return modelsStore.getModelModalitiesArray(firstModel.id);
+		return modelsStore.props.getModelModalitiesArray(firstModel.id);
 	});

 	// Ensure models are fetched when dialog opens
@@ -56,7 +56,7 @@
 	$effect(() => {
 		if (open && isRouter && modelId) {
 			isLoadingRouterProps = true;
-			modelsStore
+			modelsStore.props
 				.fetchModelProps(modelId)
 				.then((props) => {
 					routerModelProps = props;
@@ -14,7 +14,9 @@

 	let mcpServers = $derived(mcpStore.getServers().filter((s) => s.enabled));
 	let enabledMcpServersForChat = $derived(
-		mcpServers.filter((s) => conversationsStore.isMcpServerEnabledForChat(s.id) && s.url.trim())
+		mcpServers.filter(
+			(s) => conversationsStore.preferences.isMcpServerEnabledForChat(s.id) && s.url.trim()
+		)
 	);
 	let healthyEnabledMcpServers = $derived(
 		enabledMcpServersForChat.filter((s) => {
@@ -2,7 +2,7 @@
 	import McpResourcesBrowserEmptyState from './McpResourcesBrowserEmptyState.svelte';
 	import McpResourcesBrowserHeader from './McpResourcesBrowserHeader.svelte';
 	import McpResourcesBrowserServerItem from './McpResourcesBrowserServerItem.svelte';
-	import { mcpResourceStore, mcpStore } from '$lib/stores';
+	import { mcpStore } from '$lib/stores';
 	import type { MCPResourceInfo, MCPResourceTemplateInfo, MCPServerResources } from '$lib/types';
 	import { parseResourcePath } from '$lib/utils';
 	import { SvelteMap, SvelteSet } from 'svelte/reactivity';
@@ -31,8 +31,8 @@
 	let expandedFolders = new SvelteSet<string>();
 	let searchQuery = $state('');

-	const resources = $derived(mcpResourceStore.serverResources);
-	const isLoading = $derived(mcpResourceStore.isLoading);
+	const resources = $derived(mcpStore.resources.serverResources);
+	const isLoading = $derived(mcpStore.resources.isLoading);

 	const filteredResources = $derived.by(() => {
 		if (!searchQuery.trim()) {
@@ -116,7 +116,7 @@

 		if (status === ServerModelStatus.LOADING) return;

-		await modelsStore.unloadModel(modelId);
+		await modelsStore.status.unload(modelId);
 	}

 	export function open() {
@@ -174,9 +174,9 @@
 		{@const triggerLoading =
 			!!triggerModel &&
 			(triggerStatus === ServerModelStatus.LOADING ||
-				modelsStore.isModelOperationInProgress(triggerModel))}
+				modelsStore.status.isOperationInProgress(triggerModel))}
 		{@const triggerLoadPercent = triggerLoading
-			? Math.round(modelLoadFraction(modelsStore.getLoadProgress(triggerModel)) * 100)
+			? Math.round(modelLoadFraction(modelsStore.status.getLoadProgress(triggerModel)) * 100)
 			: 0}

 		{#if ms.isRouter}
@@ -47,7 +47,7 @@

 		return (model?.status?.value as ServerModelStatus) ?? null;
 	});
-	let isOperationInProgress = $derived(modelsStore.isModelOperationInProgress(option.model));
+	let isOperationInProgress = $derived(modelsStore.status.isOperationInProgress(option.model));
 	let isFailed = $derived(serverStatus === ServerModelStatus.FAILED);
 	let isSleeping = $derived(serverStatus === ServerModelStatus.SLEEPING);
 	let isLoaded = $derived(
@@ -55,7 +55,7 @@
 	);
 	let isLoading = $derived(serverStatus === ServerModelStatus.LOADING || isOperationInProgress);

-	let loadProgress = $derived(isLoading ? modelsStore.getLoadProgress(option.model) : null);
+	let loadProgress = $derived(isLoading ? modelsStore.status.getLoadProgress(option.model) : null);
 	let loadPercent = $derived(Math.round(modelLoadFraction(loadProgress) * 100));
 	let loadTitle = $derived(modelLoadProgressText(loadProgress));
 </script>
@@ -138,7 +138,7 @@
 						icon={RotateCw}
 						tooltip="Retry loading model"
 						class="h-3 w-3 text-red-500 hover:text-foreground"
-						onclick={() => modelsStore.loadModel(option.model)}
+						onclick={() => modelsStore.status.load(option.model)}
 						stopPropagationOnClick
 					/>
 				</div>
@@ -157,7 +157,7 @@
 						class="h-3 w-3 text-red-500 hover:text-red-600 [@media(pointer:coarse)]:text-amber-500 [@media(pointer:coarse)]:hover:text-amber-600"
 						onclick={(e) => {
 							e?.stopPropagation();
-							modelsStore.unloadModel(option.model);
+							modelsStore.status.unload(option.model);
 						}}
 					/>
 				</div>
@@ -174,7 +174,7 @@
 						icon={PowerOff}
 						tooltip="Unload model"
 						class="h-3 w-3 text-red-500 hover:text-red-600 [@media(pointer:coarse)]:text-green-500 [@media(pointer:coarse)]:hover:text-green-600"
-						onclick={() => modelsStore.unloadModel(option.model)}
+						onclick={() => modelsStore.status.unload(option.model)}
 						stopPropagationOnClick
 					/>
 				</div>
@@ -191,7 +191,7 @@
 						icon={Power}
 						tooltip="Load model"
 						class="h-3 w-3 [@media(pointer:coarse)]:text-muted-foreground"
-						onclick={() => modelsStore.loadModel(option.model)}
+						onclick={() => modelsStore.status.load(option.model)}
 						stopPropagationOnClick
 					/>
 				</div>
@@ -72,9 +72,9 @@
 		{@const triggerLoading =
 			!!triggerModel &&
 			(triggerStatus === ServerModelStatus.LOADING ||
-				modelsStore.isModelOperationInProgress(triggerModel))}
+				modelsStore.status.isOperationInProgress(triggerModel))}
 		{@const triggerLoadPercent = triggerLoading
-			? Math.round(modelLoadFraction(modelsStore.getLoadProgress(triggerModel)) * 100)
+			? Math.round(modelLoadFraction(modelsStore.status.getLoadProgress(triggerModel)) * 100)
 			: 0}

 		{#if ms.isRouter}
@@ -15,7 +15,7 @@
 		NUMERIC_FIELDS,
 		POSITIVE_INTEGER_FIELDS,
 		SETTINGS_CHAT_SECTIONS,
-		SETTINGS_SECTION_TITLES
+		SETTINGS_SECTION_SLUGS
 	} from '$lib/constants';
 	import { ColorMode } from '$lib/enums/ui.enums';
 	import { RouterService } from '$lib/services/router.service';
@@ -46,13 +46,13 @@
 	let fetchInitiated = false;

 	$effect(() => {
-		if (serverStore.isRouterMode && currentSection.fields && !fetchInitiated) {
+		if (serverStore.isRouterMode && currentSection.fields?.length && !fetchInitiated) {
 			fetchInitiated = true;

 			void modelsStore
 				.fetch()
 				.then(() => modelsStore.fetchRouterModels())
-				.then(() => modelsStore.fetchModalitiesForLoadedModels())
+				.then(() => modelsStore.props.fetchModalitiesForLoadedModels())
 				.then(() => modelsStore.ensureFirstModelSelected());
 		}
 	});
@@ -148,9 +148,9 @@
 						<h3 class="text-lg font-semibold">{currentSection.title}</h3>
 					</div>

-					{#if currentSection.title === SETTINGS_SECTION_TITLES.TOOLS}
+					{#if currentSection.slug === SETTINGS_SECTION_SLUGS.TOOLS}
 						<SettingsChatToolsTab />
-					{:else if currentSection.title === SETTINGS_SECTION_TITLES.IMPORT_EXPORT}
+					{:else if currentSection.slug === SETTINGS_SECTION_SLUGS.IMPORT_EXPORT}
 						<SettingsChatImportExportTab />
 					{:else if currentSection.fields}
 						<div class="space-y-6">
@@ -161,7 +161,7 @@
 								onThemeChange={handleThemeChange}
 							/>

-							{#if currentSection.title === SETTINGS_SECTION_TITLES.GENERAL}
+							{#if currentSection.slug === SETTINGS_SECTION_SLUGS.GENERAL}
 								<div class="flex justify-end">
 									<Button variant="outline" onclick={() => window.location.reload()}>
 										<RefreshCw class="h-3 w-3" />
@@ -23,13 +23,13 @@
 	let { fields, localConfig, onConfigChange, onThemeChange }: Props = $props();

 	let currentModelParams = $derived.by(() => {
-		void modelsStore.propsCacheVersion;
+		void modelsStore.props.cacheVersion;

 		if (serverStore.isRouterMode) {
 			const currentModelName = modelsStore.selectedModelName;

 			if (currentModelName) {
-				const currentModelProps = modelsStore.getModelProps(currentModelName);
+				const currentModelProps = modelsStore.props.getModelProps(currentModelName);

 				return (currentModelProps?.default_generation_settings?.params ?? {}) as Record<
 					string,
@@ -121,11 +121,13 @@
 				{:else}
 					<McpServerCard
 						{server}
-						enabled={conversationsStore.isMcpServerEnabledForChat(server.id)}
+						enabled={conversationsStore.preferences.isMcpServerEnabledForChat(server.id)}
 						onToggle={async () => {
-							const wasEnabled = conversationsStore.isMcpServerEnabledForChat(server.id);
+							const wasEnabled = conversationsStore.preferences.isMcpServerEnabledForChat(
+								server.id
+							);

-							await conversationsStore.toggleMcpServerForChat(server.id);
+							await conversationsStore.preferences.toggleMcpServerForChat(server.id);

 							if (!wasEnabled) {
 								// Promote the connection so tools/prompts/resources become
@@ -74,7 +74,7 @@ export const ATTACHMENT_PROMPT_ITEMS: AttachmentMenuItem[] = [
 		enabledWhen: AttachmentItemEnabledWhen.ALWAYS,
 		icon: Zap,
 		id: AttachmentMenuItemId.MCP_PROMPT,
-		label: 'MCP Prompt',
+		label: 'MCP Prompts',
 		visibleWhen: AttachmentItemVisibleWhen.HAS_MCP_PROMPTS_SUPPORT
 	}
 ];
@@ -32,13 +32,3 @@ export const MCP_RESOURCE_CACHE = {
 	/** TTL for MCP resource cache entries in milliseconds (5 minutes) */
 	TTL_MS: 5 * 60 * 1000
 } as const;
-
-/**
- * Limits for pruning inactive conversation states held in memory.
- */
-export const INACTIVE_CONVERSATION = {
-	/** Maximum age (in ms) for inactive conversation states before cleanup (30 minutes) */
-	MAX_AGE_MS: 30 * 60 * 1000,
-	/** Maximum number of inactive conversation states to keep in memory */
-	MAX_STATES: 10
-} as const;
@@ -48,7 +48,7 @@ export * from './pwa.constants';
 export * from './routes.constants';
 export * from './sandbox.constants';
 export * from './settings-keys.constants';
-export * from './settings-registry.constants';
+export * from './settings.constants';
 export * from './special-characters.constants';
 export * from './stream.constants';
 export * from './supported-file-types.constants';
@@ -10,18 +10,6 @@ export const URL_PARAMS = {
 	QUERY: 'q'
 } as const;

-/** Settings section slugs — used for routes and navigation. */
-export const SETTINGS_SECTION_SLUGS = {
-	AGENTIC: 'agentic',
-	DEVELOPER: 'developer',
-	DISPLAY: 'display',
-	GENERAL: 'general',
-	IMPORT_EXPORT: 'import-export',
-	PENALTIES: 'penalties',
-	SAMPLING: 'sampling',
-	TOOLS: 'tools'
-} as const;
-
 export const ROUTES = {
 	/** Chat base — for dynamic chat URLs use RouterService. */
 	CHAT: '#/chat',
@@ -33,6 +21,8 @@ export const ROUTES = {
 	SEARCH: '#/search',
 	/** Settings base — for dynamic settings URLs use RouterService. */
 	SETTINGS: '#/settings',
+	/** Exit destination for the settings view (fallback when no referrer). */
+	SETTINGS_EXIT: '#/',
 	/** Root — start of the app. */
 	START: '#/'
 } as const;
@@ -1,3 +1,5 @@
+import { UrlProtocol } from '$lib/enums';
+
 const STD = ['com', 'net', 'org', 'gov', 'edu'] as const;
 const STD_MIL = [...STD, 'mil'] as const;
 const ccTLD_PREFIXES: Record<string, readonly string[]> = {
@@ -184,3 +186,7 @@ export const WILDCARD_PUBLIC_SUFFIXES = buildSuffixSet(WILDCARD_BASES);

 // Matches one or more trailing "/" characters at the end of a URL/path.
 export const TRAILING_SLASHES_REGEX = /\/+$/;
+
+// Protocols that apiFetch treats as absolute and passes through untouched.
+// Add a protocol here when a caller needs to fetch an absolute URL with it.
+export const API_ABSOLUTE_URL_PROTOCOLS = [UrlProtocol.HTTP, UrlProtocol.HTTPS] as const;
@@ -14,18 +14,14 @@ export interface AutoScrollOptions {
 */
 export class AutoScrollController {
 	private _autoScrollEnabled = $state(true);
-	private _userScrolledUp = $state(false);
-	private _lastScrollTop = $state(0);
-	private _scrollInterval: ReturnType<typeof setInterval> | undefined;
 	private _container: HTMLElement | undefined;
 	private _disabled: boolean;
+	private _lastScrollTop = $state(0);
 	private _mutationObserver: MutationObserver | null = null;
-	private _rafPending = false;
 	private _observerEnabled = false;
-	constructor(options: AutoScrollOptions = {}) {
-		this._disabled = options.disabled ?? false;
-	}
-
+	private _rafPending = false;
+	private _scrollInterval: ReturnType<typeof setInterval> | undefined;
+	private _userScrolledUp = $state(false);
 	get autoScrollEnabled(): boolean {
 		return this._autoScrollEnabled;
 	}
@@ -34,6 +30,71 @@ export class AutoScrollController {
 		return this._userScrolledUp;
 	}

+	constructor(options: AutoScrollOptions = {}) {
+		this._disabled = options.disabled ?? false;
+	}
+
+	/**
+	 * Cleans up resources. Call this in onDestroy or when the component unmounts.
+	 */
+	destroy(): void {
+		this.stopInterval();
+		this._doStopObserving();
+	}
+
+	/**
+	 * Enables auto-scroll (e.g., when user sends a message).
+	 */
+	enable(): void {
+		if (this._disabled) return;
+
+		this._userScrolledUp = false;
+		this._autoScrollEnabled = true;
+	}
+
+	/**
+	 * Handles scroll events to detect user scroll direction and toggle auto-scroll.
+	 */
+	handleScroll(): void {
+		if (this._disabled || !this._container) return;
+
+		const { clientHeight, scrollHeight, scrollTop } = this._container;
+		const distanceFromBottom = scrollHeight - clientHeight - scrollTop;
+		const isScrollingUp = scrollTop < this._lastScrollTop;
+		const isAtBottom = distanceFromBottom < AUTO_SCROLL_AT_BOTTOM_THRESHOLD;
+
+		if (isScrollingUp && !isAtBottom) {
+			this._userScrolledUp = true;
+			this._autoScrollEnabled = false;
+		} else if (isAtBottom && this._userScrolledUp) {
+			this._userScrolledUp = false;
+			this._autoScrollEnabled = true;
+		}
+
+		this._lastScrollTop = scrollTop;
+	}
+
+	/**
+	 * Resets scroll state when switching conversations.
+	 */
+	resetScrollState(): void {
+		this._userScrolledUp = false;
+		this._autoScrollEnabled = !this._disabled;
+
+		if (this._container) {
+			this._lastScrollTop = this._container.scrollTop;
+		}
+	}
+
+	/**
+	 * Scrolls the container to the bottom instantly.
+	 */
+	scrollToBottom(): void {
+		if (this._disabled || !this._container) return;
+
+		this._container.scrollTop = this._container.scrollHeight;
+	}
+
 	/**
 	 * Binds the controller to a scrollable container element.
 	 */
@@ -63,59 +124,6 @@ export class AutoScrollController {
 		}
 	}

-	/**
-	 * Handles scroll events to detect user scroll direction and toggle auto-scroll.
-	 */
-	handleScroll(): void {
-		if (this._disabled || !this._container) return;
-
-		const { clientHeight, scrollHeight, scrollTop } = this._container;
-		const distanceFromBottom = scrollHeight - clientHeight - scrollTop;
-		const isScrollingUp = scrollTop < this._lastScrollTop;
-		const isAtBottom = distanceFromBottom < AUTO_SCROLL_AT_BOTTOM_THRESHOLD;
-
-		if (isScrollingUp && !isAtBottom) {
-			this._userScrolledUp = true;
-			this._autoScrollEnabled = false;
-		} else if (isAtBottom && this._userScrolledUp) {
-			this._userScrolledUp = false;
-			this._autoScrollEnabled = true;
-		}
-
-		this._lastScrollTop = scrollTop;
-	}
-
-	/**
-	 * Scrolls the container to the bottom instantly.
-	 */
-	scrollToBottom(): void {
-		if (this._disabled || !this._container) return;
-
-		this._container.scrollTop = this._container.scrollHeight;
-	}
-
-	/**
-	 * Enables auto-scroll (e.g., when user sends a message).
-	 */
-	enable(): void {
-		if (this._disabled) return;
-
-		this._userScrolledUp = false;
-		this._autoScrollEnabled = true;
-	}
-
-	/**
-	 * Resets scroll state when switching conversations.
-	 */
-	resetScrollState(): void {
-		this._userScrolledUp = false;
-		this._autoScrollEnabled = !this._disabled;
-
-		if (this._container) {
-			this._lastScrollTop = this._container.scrollTop;
-		}
-	}
-
 	/**
 	 * Starts the auto-scroll interval for continuous scrolling during streaming.
 	 */
@@ -127,6 +135,18 @@ export class AutoScrollController {
 		}, AUTO_SCROLL_INTERVAL);
 	}

+	/**
+	 * Starts a MutationObserver on the container that auto-scrolls to bottom
+	 * on content changes. More responsive than interval-based polling.
+	 */
+	startObserving(): void {
+		this._observerEnabled = true;
+
+		if (this._container && !this._disabled && !this._mutationObserver) {
+			this._doStartObserving();
+		}
+	}
+
 	/**
 	 * Stops the auto-scroll interval.
 	 */
@@ -137,6 +157,14 @@ export class AutoScrollController {
 		}
 	}

+	/**
+	 * Stops the MutationObserver.
+	 */
+	stopObserving(): void {
+		this._observerEnabled = false;
+		this._doStopObserving();
+	}
+
 	/**
 	 * Updates the auto-scroll interval based on streaming state.
 	 * Call this in a $effect to automatically manage the interval.
@@ -157,34 +185,6 @@ export class AutoScrollController {
 		}
 	}

-	/**
-	 * Cleans up resources. Call this in onDestroy or when the component unmounts.
-	 */
-	destroy(): void {
-		this.stopInterval();
-		this._doStopObserving();
-	}
-
-	/**
-	 * Starts a MutationObserver on the container that auto-scrolls to bottom
-	 * on content changes. More responsive than interval-based polling.
-	 */
-	startObserving(): void {
-		this._observerEnabled = true;
-
-		if (this._container && !this._disabled && !this._mutationObserver) {
-			this._doStartObserving();
-		}
-	}
-
-	/**
-	 * Stops the MutationObserver.
-	 */
-	stopObserving(): void {
-		this._observerEnabled = false;
-		this._doStopObserving();
-	}
-
 	private _doStartObserving(): void {
 		if (!this._container || this._mutationObserver) return;

@@ -22,10 +22,10 @@ export function useChatScreenActiveModel() {

 	$effect(() => {
 		if (activeModelId) {
-			const cached = modelsStore.getModelProps(activeModelId);
+			const cached = modelsStore.props.getModelProps(activeModelId);

 			if (!cached) {
-				modelsStore.fetchModelProps(activeModelId).then(() => {
+				modelsStore.props.fetchModelProps(activeModelId).then(() => {
 					modelPropsVersion++;
 				});
 			}
@@ -36,7 +36,7 @@ export function useChatScreenActiveModel() {
 		if (activeModelId) {
 			void modelPropsVersion;

-			return modelsStore.modelSupportsAudio(activeModelId);
+			return modelsStore.props.modelSupportsAudio(activeModelId);
 		}

 		return false;
@@ -45,7 +45,7 @@ export function useChatScreenActiveModel() {
 		if (activeModelId) {
 			void modelPropsVersion;

-			return modelsStore.modelSupportsVideo(activeModelId);
+			return modelsStore.props.modelSupportsVideo(activeModelId);
 		}

 		return false;
@@ -54,7 +54,7 @@ export function useChatScreenActiveModel() {
 		if (activeModelId) {
 			void modelPropsVersion;

-			return modelsStore.modelSupportsVision(activeModelId);
+			return modelsStore.props.modelSupportsVision(activeModelId);
 		}

 		return false;
@@ -54,10 +54,10 @@ export function useContextGauge(): UseContextGaugeReturn {
 		const modelId = contextStatsStore.activeModelId;

 		if (modelId && contextStatsStore.isActiveModelLoaded) {
-			const cached = modelsStore.getModelProps(modelId);
+			const cached = modelsStore.props.getModelProps(modelId);

 			if (!cached) {
-				void modelsStore.fetchModelProps(modelId);
+				void modelsStore.props.fetchModelProps(modelId);
 			}
 		}
 	});
@@ -80,9 +80,9 @@ export function useContextGauge(): UseContextGaugeReturn {
 		if (!modelId || contextStatsStore.isActiveModelLoading) return;

 		try {
-			await modelsStore.loadModel(modelId);
+			await modelsStore.status.load(modelId);
 		} catch {
-			// toast already surfaced by modelsStore.loadModel
+			// toast already surfaced by modelsStore.status.load
 		}
 	}

@@ -47,7 +47,7 @@ export interface UseModelsSelectorReturn {
 export function useModelsSelector(opts: UseModelsSelectorOptions): UseModelsSelectorReturn {
 	const options = $derived(
 		modelsStore.models.filter((option) => {
-			const modelProps = modelsStore.getModelProps(option.model);
+			const modelProps = modelsStore.props.getModelProps(option.model);

 			return modelProps?.ui !== false;
 		})
@@ -103,7 +103,7 @@ export function useModelsSelector(opts: UseModelsSelectorOptions): UseModelsSele

 			if (open) {
 				modelsStore.fetchRouterModels().then(() => {
-					modelsStore.fetchModalitiesForLoadedModels();
+					modelsStore.props.fetchModalitiesForLoadedModels();
 				});
 			}

@@ -143,8 +143,8 @@ export function useModelsSelector(opts: UseModelsSelectorOptions): UseModelsSele
 		if (!onModelChange && isRouter && !modelsStore.isModelLoaded(option.model)) {
 			isLoadingModel = true;

-			modelsStore
-				.loadModel(option.model)
+			modelsStore.status
+				.load(option.model)
 				.catch((error) => console.error('Failed to load model:', error))
 				.finally(() => (isLoadingModel = false));
 		}
@@ -43,7 +43,7 @@ export function useProcessingState(): UseProcessingStateReturn {
 		}

 		// Read directly from the reactive state
-		return chatStore.activeProcessingState;
+		return chatStore.processing.activeState;
 	});

 	$effect(() => {
@@ -42,19 +42,20 @@ export function useReasoningMenu(): UseReasoningMenuReturn {
 	});
 	const modelSupportsThinking = $derived.by(() => {
 		void modelsStore.loadedModelIds;
-		void modelsStore.propsCacheVersion;
+		void modelsStore.props.cacheVersion;

 		if (serverStore.isRouterMode) {
 			const modelId = modelsStore.selectedModelName || conversationModel;

 			return (
-				modelsStore.checkModelSupportsThinking(modelId ?? '') || modelSupportsThinkingFromMessages
+				modelsStore.props.checkModelSupportsThinking(modelId ?? '') ||
+				modelSupportsThinkingFromMessages
 			);
 		}

-		return modelsStore.supportsThinking || modelSupportsThinkingFromMessages;
+		return modelsStore.props.supportsThinking || modelSupportsThinkingFromMessages;
 	});
-	const currentEffort = $derived(conversationsStore.getReasoningEffort());
+	const currentEffort = $derived(conversationsStore.preferences.getReasoningEffort());
 	const thinkingEnabled = $derived(
 		currentEffort !== ReasoningEffort.OFF && currentEffort !== ReasoningEffort.DEFAULT
 	);
@@ -76,7 +77,7 @@ export function useReasoningMenu(): UseReasoningMenuReturn {
 			return modelSupportsThinking;
 		},
 		select(level: ReasoningEffortLevel): void {
-			conversationsStore.setReasoningEffort(level.value as ReasoningEffort);
+			conversationsStore.preferences.setReasoningEffort(level.value as ReasoningEffort);
 		},
 		get thinkingEnabled() {
 			return thinkingEnabled;
@@ -35,7 +35,7 @@ export function useToolsPanel(): UseToolsPanelReturn {
 			(g) =>
 				g.source !== ToolSource.MCP ||
 				!g.serverId ||
-				conversationsStore.isMcpServerEnabledForChat(g.serverId)
+				conversationsStore.preferences.isMcpServerEnabledForChat(g.serverId)
 		)
 	);
 	const totalToolCount = $derived(activeGroups.reduce((n, g) => n + g.tools.length, 0));
@@ -73,7 +73,7 @@ export function useToolsPanel(): UseToolsPanelReturn {
 		return (
 			group.source === ToolSource.MCP &&
 			!!group.serverId &&
-			!conversationsStore.isMcpServerEnabledForChat(group.serverId)
+			!conversationsStore.preferences.isMcpServerEnabledForChat(group.serverId)
 		);
 	}

@@ -16,187 +16,6 @@ import {
 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)
@@ -262,6 +81,171 @@ export class ConversationTransferService {
 		ConversationTransferService.triggerDownload(blob, archiveName);
 	}

+	/**
+	 * 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}`;
+	}
+
+	/**
+	 * 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'
+		);
+	}
+
+	/**
+	 * 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;
+	}
+
+	/**
+	 * 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);
+	}
+
+	/**
+	 * 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;
+		}
+	}
+
 	/**
 	 * Triggers a browser download of a blob under the given filename.
 	 */
@@ -1,3 +1,11 @@
+/**
+ * DatabaseService - IndexedDB persistence for conversations and messages
+ *
+ * Thin Dexie layer over the conversations/messages tables: CRUD, tree
+ * navigation (descendants, reparenting) and cascading deletes. No reactive
+ * state; consumed by conversationsStore and the chat flows.
+ */
+
 import { IDXDB_STORES, IDXDB_TABLES, STORAGE_APP_NAME } from '$lib/constants';
 import { MessageRole } from '$lib/enums';
 import type { McpServerOverride } from '$lib/types/database';
@@ -20,12 +28,99 @@ const db = new LlamaUiDatabase();

 export class DatabaseService {
 	/**
+	 * Deletes multiple conversations in a single transaction. Each deleted
+	 * conversation has its direct children reparented to the nearest surviving
+	 * ancestor (or promoted to top-level). Children also in `ids` are dropped
+	 * entirely rather than reparented.
 	 *
-	 *
-	 * Conversations
-	 *
-	 *
+	 * @param ids - Conversation IDs to delete
 	 */
+	static async bulkDeleteConversations(ids: string[]): Promise<void> {
+		const cleanIds = ids.filter((id): id is string => typeof id === 'string' && id.length > 0);
+
+		if (cleanIds.length === 0) return;
+
+		const idSet = new Set(cleanIds);
+
+		await db.transaction(
+			'rw',
+			[db[IDXDB_TABLES.conversations], db[IDXDB_TABLES.messages]],
+			async () => {
+				// Pre-load each to-delete conversation so the per-id reparent
+				// walk-up doesn't ping-pong the same ancestry chain.
+				const prefetched = new Map<string, DatabaseConversation>();
+
+				let frontier = [...cleanIds];
+
+				const requested = new Set<string>(frontier);
+
+				while (frontier.length > 0) {
+					const fetched = await db[IDXDB_TABLES.conversations].bulkGet(frontier);
+
+					frontier = [];
+					for (let i = 0; i < fetched.length; i++) {
+						const conv = fetched[i];
+
+						if (!conv || !conv.id) continue;
+
+						prefetched.set(conv.id, conv);
+						const ancestor = conv.forkedFromConversationId;
+
+						if (ancestor && !prefetched.has(ancestor) && !requested.has(ancestor)) {
+							frontier.push(ancestor);
+							requested.add(ancestor);
+						}
+					}
+				}
+
+				for (const id of cleanIds) {
+					await this.reparentDirectChildren(id, idSet, prefetched);
+				}
+
+				await db[IDXDB_TABLES.conversations].bulkDelete(cleanIds);
+				await db[IDXDB_TABLES.messages].where('convId').anyOf(cleanIds).delete();
+			}
+		);
+	}
+
+	/**
+	 * Toggles the pinned status of each conversation in `ids` inside a single
+	 * transaction. Treats `pinned === undefined` as `false`, matching the
+	 * semantics of {@link toggleConversationPin} where `!undefined` evaluates
+	 * to `true`. Returns the resulting pinned state for every id that was
+	 * updated; missing ids are omitted from the map.
+	 *
+	 * @param ids - Conversation IDs to toggle
+	 * @returns Map of id -> new pinned state
+	 */
+	static async bulkToggleConversationPins(ids: string[]): Promise<Map<string, boolean>> {
+		const cleanIds = ids.filter((id): id is string => typeof id === 'string' && id.length > 0);
+		const result = new Map<string, boolean>();
+
+		if (cleanIds.length === 0) return result;
+
+		await db.transaction('rw', db[IDXDB_TABLES.conversations], async () => {
+			const convs = await db[IDXDB_TABLES.conversations].bulkGet(cleanIds);
+			const updates: DatabaseConversation[] = [];
+
+			for (let i = 0; i < cleanIds.length; i++) {
+				const conv = convs[i];
+
+				if (!conv) continue;
+
+				const newPinned = !conv.pinned;
+
+				updates.push({ ...conv, pinned: newPinned });
+				result.set(cleanIds[i], newPinned);
+			}
+
+			if (updates.length === 0) return;
+
+			await db[IDXDB_TABLES.conversations].bulkPut(updates);
+		});
+
+		return result;
+	}

 	/**
 	 * Creates a new conversation.
@@ -51,14 +146,6 @@ export class DatabaseService {
 		return conversation;
 	}

-	/**
-	 *
-	 *
-	 * Messages
-	 *
-	 *
-	 */
-
 	/**
 	 * Creates a new message branch by adding a message and updating parent/child relationships.
 	 * Also updates the conversation's currNode to point to the new message.
@@ -96,13 +183,7 @@ export class DatabaseService {

 				// Update parent's children array if parent exists
 				if (parentId !== null) {
-					const parentMessage = await db[IDXDB_TABLES.messages].get(parentId);
-
-					if (parentMessage) {
-						await db[IDXDB_TABLES.messages].update(parentId, {
-							children: [...parentMessage.children, newMessage.id]
-						});
-					}
+					await this.addChildToParent(parentId, newMessage.id);
 				}

 				await this.updateConversation(message.convId, {
@@ -178,9 +259,7 @@ export class DatabaseService {
 			};

 			await db[IDXDB_TABLES.messages].add(systemMessage);
-			await db[IDXDB_TABLES.messages].update(parentId, {
-				children: [...parentMessage.children, systemMessage.id]
-			});
+			await this.addChildToParent(parentId, systemMessage.id);

 			return systemMessage;
 		});
@@ -230,121 +309,6 @@ export class DatabaseService {
 		);
 	}

-	/**
-	 * Reparents direct children of `parentId` to the nearest surviving
-	 * ancestor (or promotes them to top-level when the immediate parent was
-	 * top-level). Walking skips any ancestor listed in `excludeIds`, since
-	 * those will be deleted in the same batch — leaving a grandchild pointing
-	 * at an `excludeIds` entry would orphan it. Children whose own id is in
-	 * `excludeIds` are dropped from the updates (the bulk-delete pass will
-	 * remove them). `prefetched` may carry a pre-fetched ancestor map to
-	 * avoid repeat reads inside a bulk transaction.
-	 */
-	private static async reparentDirectChildren(
-		parentId: string,
-		excludeIds: ReadonlySet<string> = new Set(),
-		prefetched?: ReadonlyMap<string, DatabaseConversation>
-	): Promise<void> {
-		const conv = prefetched?.get(parentId) ?? (await db[IDXDB_TABLES.conversations].get(parentId));
-
-		if (!conv) return;
-
-		let newParent = conv.forkedFromConversationId;
-
-		const visited = new Set<string>([parentId]);
-
-		while (newParent && excludeIds.has(newParent)) {
-			if (visited.has(newParent)) {
-				newParent = undefined;
-
-				break;
-			}
-
-			visited.add(newParent);
-			const next =
-				prefetched?.get(newParent) ?? (await db[IDXDB_TABLES.conversations].get(newParent));
-
-			if (!next) {
-				newParent = undefined;
-
-				break;
-			}
-
-			newParent = next.forkedFromConversationId;
-		}
-
-		const directChildren = await db[IDXDB_TABLES.conversations]
-			.filter((c) => c.forkedFromConversationId === parentId)
-			.toArray();
-		const updates: DatabaseConversation[] = [];
-
-		for (const child of directChildren) {
-			if (excludeIds.has(child.id)) continue;
-
-			updates.push({ ...child, forkedFromConversationId: newParent });
-		}
-
-		if (updates.length === 0) return;
-
-		await db[IDXDB_TABLES.conversations].bulkPut(updates);
-	}
-
-	/**
-	 * Deletes multiple conversations in a single transaction. Each deleted
-	 * conversation has its direct children reparented to the nearest surviving
-	 * ancestor (or promoted to top-level). Children also in `ids` are dropped
-	 * entirely rather than reparented.
-	 *
-	 * @param ids - Conversation IDs to delete
-	 */
-	static async bulkDeleteConversations(ids: string[]): Promise<void> {
-		const cleanIds = ids.filter((id): id is string => typeof id === 'string' && id.length > 0);
-
-		if (cleanIds.length === 0) return;
-
-		const idSet = new Set(cleanIds);
-
-		await db.transaction(
-			'rw',
-			[db[IDXDB_TABLES.conversations], db[IDXDB_TABLES.messages]],
-			async () => {
-				// Pre-load each to-delete conversation so the per-id reparent
-				// walk-up doesn't ping-pong the same ancestry chain.
-				const prefetched = new Map<string, DatabaseConversation>();
-
-				let frontier = [...cleanIds];
-
-				const requested = new Set<string>(frontier);
-
-				while (frontier.length > 0) {
-					const fetched = await db[IDXDB_TABLES.conversations].bulkGet(frontier);
-
-					frontier = [];
-					for (let i = 0; i < fetched.length; i++) {
-						const conv = fetched[i];
-
-						if (!conv || !conv.id) continue;
-
-						prefetched.set(conv.id, conv);
-						const ancestor = conv.forkedFromConversationId;
-
-						if (ancestor && !prefetched.has(ancestor) && !requested.has(ancestor)) {
-							frontier.push(ancestor);
-							requested.add(ancestor);
-						}
-					}
-				}
-
-				for (const id of cleanIds) {
-					await this.reparentDirectChildren(id, idSet, prefetched);
-				}
-
-				await db[IDXDB_TABLES.conversations].bulkDelete(cleanIds);
-				await db[IDXDB_TABLES.messages].where('convId').anyOf(cleanIds).delete();
-			}
-		);
-	}
-
 	/**
 	 * Deletes a message and removes it from its parent's children array.
 	 *
@@ -356,17 +320,8 @@ export class DatabaseService {

 			if (!message) return;

-			// Remove this message from its parent's children array
-			if (message.parent) {
-				const parent = await db[IDXDB_TABLES.messages].get(message.parent);
+			await this.removeChildFromParent(messageId);

-				if (parent) {
-					parent.children = parent.children.filter((childId: string) => childId !== messageId);
-					await db[IDXDB_TABLES.messages].put(parent);
-				}
-			}
-
-			// Delete the message
 			await db[IDXDB_TABLES.messages].delete(messageId);
 		});
 	}
@@ -389,20 +344,10 @@ export class DatabaseService {
 				.where('convId')
 				.equals(conversationId)
 				.toArray();
-			// Find all descendant messages
 			const descendants = findDescendantMessages(allMessages, messageId);
 			const allToDelete = [messageId, ...descendants];
-			// Get the message to delete for parent cleanup
-			const message = await db[IDXDB_TABLES.messages].get(messageId);

-			if (message && message.parent) {
-				const parent = await db[IDXDB_TABLES.messages].get(message.parent);
-
-				if (parent) {
-					parent.children = parent.children.filter((childId: string) => childId !== messageId);
-					await db[IDXDB_TABLES.messages].put(parent);
-				}
-			}
+			await this.removeChildFromParent(messageId);

 			// Delete all messages in the branch
 			await db[IDXDB_TABLES.messages].bulkDelete(allToDelete);
@@ -411,243 +356,6 @@ export class DatabaseService {
 		});
 	}

-	/**
-	 * Gets all conversations, sorted by last modified time (newest first).
-	 *
-	 * @returns Array of conversations
-	 */
-	static async getAllConversations(): Promise<DatabaseConversation[]> {
-		return await db[IDXDB_TABLES.conversations].orderBy('lastModified').reverse().toArray();
-	}
-
-	/**
-	 * Gets a conversation by ID.
-	 *
-	 * @param id - Conversation ID
-	 * @returns The conversation if found, otherwise undefined
-	 */
-	static async getConversation(id: string): Promise<DatabaseConversation | undefined> {
-		return await db[IDXDB_TABLES.conversations].get(id);
-	}
-
-	/**
-	 * Gets all messages in a conversation, sorted by timestamp (oldest first).
-	 *
-	 * @param convId - Conversation ID
-	 * @returns Array of messages in the conversation
-	 */
-	static async getConversationMessages(convId: string): Promise<DatabaseMessage[]> {
-		return await db[IDXDB_TABLES.messages].where('convId').equals(convId).sortBy('timestamp');
-	}
-
-	/**
-	 * Loads multiple conversations with all of their messages in two bulk
-	 * reads. Missing conversations are silently omitted from the result.
-	 *
-	 * @param convIds - Conversation IDs to load
-	 * @returns Map of id -> { conv, messages }. Messages are sorted ascending by timestamp.
-	 */
-	static async getConversationsWithMessages(
-		convIds: string[]
-	): Promise<Map<string, ExportedConversation>> {
-		const result = new Map<string, ExportedConversation>();
-		const cleanIds = convIds.filter((id): id is string => typeof id === 'string' && id.length > 0);
-
-		if (cleanIds.length === 0) return result;
-
-		const [convs, allMessages] = await Promise.all([
-			db[IDXDB_TABLES.conversations].bulkGet(cleanIds),
-			db[IDXDB_TABLES.messages].where('convId').anyOf(cleanIds).toArray()
-		]);
-		const messagesByConv = new Map<string, DatabaseMessage[]>();
-
-		for (const msg of allMessages) {
-			const bucket = messagesByConv.get(msg.convId);
-
-			if (bucket) bucket.push(msg);
-			else messagesByConv.set(msg.convId, [msg]);
-		}
-
-		for (let i = 0; i < cleanIds.length; i++) {
-			const conv = convs[i];
-
-			if (!conv) continue;
-
-			const messages = (messagesByConv.get(conv.id) ?? []).sort(
-				(a, b) => a.timestamp - b.timestamp
-			);
-
-			result.set(conv.id, { conv, messages });
-		}
-
-		return result;
-	}
-
-	/**
-	 * Updates a conversation. `lastModified` is never stamped implicitly;
-	 * pass it in `updates` to bump the conversation in recency ordering.
-	 *
-	 * @param id - Conversation ID
-	 * @param updates - Partial updates to apply
-	 * @returns Promise that resolves when the conversation is updated
-	 */
-	static async updateConversation(
-		id: string,
-		updates: Partial<Omit<DatabaseConversation, 'id'>>
-	): Promise<void> {
-		await db[IDXDB_TABLES.conversations].update(id, updates);
-	}
-
-	/**
-	 *
-	 *
-	 * Navigation
-	 *
-	 *
-	 */
-
-	/**
-	 * Toggles the pinned status of a conversation.
-	 *
-	 * @param id - Conversation ID
-	 * @returns The new pinned status
-	 */
-	static async toggleConversationPin(id: string): Promise<boolean> {
-		const conversation = await db[IDXDB_TABLES.conversations].get(id);
-
-		if (!conversation) {
-			throw new Error(`Conversation ${id} not found`);
-		}
-
-		const newPinnedState = !conversation.pinned;
-
-		await this.updateConversation(id, { pinned: newPinnedState });
-
-		return newPinnedState;
-	}
-
-	/**
-	 * Toggles the pinned status of each conversation in `ids` inside a single
-	 * transaction. Treats `pinned === undefined` as `false`, matching the
-	 * semantics of {@link toggleConversationPin} where `!undefined` evaluates
-	 * to `true`. Returns the resulting pinned state for every id that was
-	 * updated; missing ids are omitted from the map.
-	 *
-	 * @param ids - Conversation IDs to toggle
-	 * @returns Map of id -> new pinned state
-	 */
-	static async bulkToggleConversationPins(ids: string[]): Promise<Map<string, boolean>> {
-		const cleanIds = ids.filter((id): id is string => typeof id === 'string' && id.length > 0);
-		const result = new Map<string, boolean>();
-
-		if (cleanIds.length === 0) return result;
-
-		await db.transaction('rw', db[IDXDB_TABLES.conversations], async () => {
-			const convs = await db[IDXDB_TABLES.conversations].bulkGet(cleanIds);
-			const updates: DatabaseConversation[] = [];
-
-			for (let i = 0; i < cleanIds.length; i++) {
-				const conv = convs[i];
-
-				if (!conv) continue;
-
-				const newPinned = !conv.pinned;
-
-				updates.push({ ...conv, pinned: newPinned });
-				result.set(cleanIds[i], newPinned);
-			}
-
-			if (updates.length === 0) return;
-
-			await db[IDXDB_TABLES.conversations].bulkPut(updates);
-		});
-
-		return result;
-	}
-
-	/**
-	 * Updates the conversation's current node (active branch).
-	 * This determines which conversation path is currently being viewed.
-	 *
-	 * @param convId - Conversation ID
-	 * @param nodeId - Message ID to set as current node
-	 */
-	static async updateCurrentNode(convId: string, nodeId: string): Promise<void> {
-		await this.updateConversation(convId, {
-			currNode: nodeId
-		});
-	}
-
-	/**
-	 * Updates a message.
-	 *
-	 * @param id - Message ID
-	 * @param updates - Partial updates to apply
-	 * @returns Promise that resolves when the message is updated
-	 */
-	static async updateMessage(
-		id: string,
-		updates: Partial<Omit<DatabaseMessage, 'id'>>
-	): Promise<void> {
-		await db[IDXDB_TABLES.messages].update(id, updates);
-	}
-
-	/**
-	 *
-	 *
-	 * Import
-	 *
-	 *
-	 */
-
-	/**
-	 * Imports multiple conversations and their messages.
-	 * Skips conversations that already exist.
-	 *
-	 * @param data - Array of { conv, messages } objects
-	 * @returns The conversations written to the database and the ones skipped
-	 */
-	static async importConversations(
-		data: { conv: DatabaseConversation; messages: DatabaseMessage[] }[]
-	): Promise<{ imported: DatabaseConversation[]; skipped: DatabaseConversation[] }> {
-		const imported: DatabaseConversation[] = [];
-		const skipped: DatabaseConversation[] = [];
-
-		return await db.transaction(
-			'rw',
-			[db[IDXDB_TABLES.conversations], db[IDXDB_TABLES.messages]],
-			async () => {
-				for (const item of data) {
-					const { conv, messages } = item;
-					const existing = await db[IDXDB_TABLES.conversations].get(conv.id);
-
-					if (existing) {
-						skipped.push(conv);
-
-						continue;
-					}
-
-					await db[IDXDB_TABLES.conversations].add(conv);
-					for (const msg of messages) {
-						await db[IDXDB_TABLES.messages].put(msg);
-					}
-
-					imported.push(conv);
-				}
-
-				return { imported, skipped };
-			}
-		);
-	}
-
-	/**
-	 *
-	 *
-	 * Forking
-	 *
-	 *
-	 */
-
 	/**
 	 * Forks a conversation at a specific message, creating a new conversation
 	 * containing all messages from the root up to (and including) the target message.
@@ -726,13 +434,272 @@ export class DatabaseService {
 				};

 				await db[IDXDB_TABLES.conversations].add(newConv);
-
-				for (const msg of clonedMessages) {
-					await db[IDXDB_TABLES.messages].add(msg);
-				}
+				await db[IDXDB_TABLES.messages].bulkAdd(clonedMessages);

 				return newConv;
 			}
 		);
 	}
+
+	/**
+	 * Gets all conversations, sorted by last modified time (newest first).
+	 *
+	 * @returns Array of conversations
+	 */
+	static async getAllConversations(): Promise<DatabaseConversation[]> {
+		return await db[IDXDB_TABLES.conversations].orderBy('lastModified').reverse().toArray();
+	}
+
+	/**
+	 * Gets a conversation by ID.
+	 *
+	 * @param id - Conversation ID
+	 * @returns The conversation if found, otherwise undefined
+	 */
+	static async getConversation(id: string): Promise<DatabaseConversation | undefined> {
+		return await db[IDXDB_TABLES.conversations].get(id);
+	}
+
+	/**
+	 * Gets all messages in a conversation, sorted by timestamp (oldest first).
+	 *
+	 * @param convId - Conversation ID
+	 * @returns Array of messages in the conversation
+	 */
+	static async getConversationMessages(convId: string): Promise<DatabaseMessage[]> {
+		return await db[IDXDB_TABLES.messages].where('convId').equals(convId).sortBy('timestamp');
+	}
+
+	/**
+	 * Loads multiple conversations with all of their messages in two bulk
+	 * reads. Missing conversations are silently omitted from the result.
+	 *
+	 * @param convIds - Conversation IDs to load
+	 * @returns Map of id -> { conv, messages }. Messages are sorted ascending by timestamp.
+	 */
+	static async getConversationsWithMessages(
+		convIds: string[]
+	): Promise<Map<string, ExportedConversation>> {
+		const result = new Map<string, ExportedConversation>();
+		const cleanIds = convIds.filter((id): id is string => typeof id === 'string' && id.length > 0);
+
+		if (cleanIds.length === 0) return result;
+
+		const [convs, allMessages] = await Promise.all([
+			db[IDXDB_TABLES.conversations].bulkGet(cleanIds),
+			db[IDXDB_TABLES.messages].where('convId').anyOf(cleanIds).toArray()
+		]);
+		const messagesByConv = new Map<string, DatabaseMessage[]>();
+
+		for (const msg of allMessages) {
+			const bucket = messagesByConv.get(msg.convId);
+
+			if (bucket) bucket.push(msg);
+			else messagesByConv.set(msg.convId, [msg]);
+		}
+
+		for (let i = 0; i < cleanIds.length; i++) {
+			const conv = convs[i];
+
+			if (!conv) continue;
+
+			const messages = (messagesByConv.get(conv.id) ?? []).sort(
+				(a, b) => a.timestamp - b.timestamp
+			);
+
+			result.set(conv.id, { conv, messages });
+		}
+
+		return result;
+	}
+
+	/**
+	 * Imports multiple conversations and their messages.
+	 * Skips conversations that already exist.
+	 *
+	 * @param data - Array of { conv, messages } objects
+	 * @returns The conversations written to the database and the ones skipped
+	 */
+	static async importConversations(
+		data: { conv: DatabaseConversation; messages: DatabaseMessage[] }[]
+	): Promise<{ imported: DatabaseConversation[]; skipped: DatabaseConversation[] }> {
+		const imported: DatabaseConversation[] = [];
+		const skipped: DatabaseConversation[] = [];
+
+		return await db.transaction(
+			'rw',
+			[db[IDXDB_TABLES.conversations], db[IDXDB_TABLES.messages]],
+			async () => {
+				for (const item of data) {
+					const { conv, messages } = item;
+					const existing = await db[IDXDB_TABLES.conversations].get(conv.id);
+
+					if (existing) {
+						skipped.push(conv);
+
+						continue;
+					}
+
+					await db[IDXDB_TABLES.conversations].add(conv);
+					for (const msg of messages) {
+						await db[IDXDB_TABLES.messages].put(msg);
+					}
+
+					imported.push(conv);
+				}
+
+				return { imported, skipped };
+			}
+		);
+	}
+
+	/**
+	 * Toggles the pinned status of a conversation.
+	 *
+	 * @param id - Conversation ID
+	 * @returns The new pinned status
+	 */
+	static async toggleConversationPin(id: string): Promise<boolean> {
+		const conversation = await db[IDXDB_TABLES.conversations].get(id);
+
+		if (!conversation) {
+			throw new Error(`Conversation ${id} not found`);
+		}
+
+		const newPinnedState = !conversation.pinned;
+
+		await this.updateConversation(id, { pinned: newPinnedState });
+
+		return newPinnedState;
+	}
+
+	/**
+	 * Updates a conversation. `lastModified` is never stamped implicitly;
+	 * pass it in `updates` to bump the conversation in recency ordering.
+	 *
+	 * @param id - Conversation ID
+	 * @param updates - Partial updates to apply
+	 * @returns Promise that resolves when the conversation is updated
+	 */
+	static async updateConversation(
+		id: string,
+		updates: Partial<Omit<DatabaseConversation, 'id'>>
+	): Promise<void> {
+		await db[IDXDB_TABLES.conversations].update(id, updates);
+	}
+
+	/**
+	 * Updates the conversation's current node (active branch).
+	 * This determines which conversation path is currently being viewed.
+	 *
+	 * @param convId - Conversation ID
+	 * @param nodeId - Message ID to set as current node
+	 */
+	static async updateCurrentNode(convId: string, nodeId: string): Promise<void> {
+		await this.updateConversation(convId, {
+			currNode: nodeId
+		});
+	}
+
+	/**
+	 * Updates a message.
+	 *
+	 * @param id - Message ID
+	 * @param updates - Partial updates to apply
+	 * @returns Promise that resolves when the message is updated
+	 */
+	static async updateMessage(
+		id: string,
+		updates: Partial<Omit<DatabaseMessage, 'id'>>
+	): Promise<void> {
+		await db[IDXDB_TABLES.messages].update(id, updates);
+	}
+
+	/**
+	 * Appends a child id to a parent message's children array.
+	 */
+	private static async addChildToParent(parentId: string, childId: string): Promise<void> {
+		const parent = await db[IDXDB_TABLES.messages].get(parentId);
+
+		if (!parent) return;
+
+		await db[IDXDB_TABLES.messages].update(parentId, {
+			children: [...parent.children, childId]
+		});
+	}
+
+	/**
+	 * Removes a child id from its parent message's children array.
+	 */
+	private static async removeChildFromParent(messageId: string): Promise<void> {
+		const message = await db[IDXDB_TABLES.messages].get(messageId);
+
+		if (!message?.parent) return;
+
+		const parent = await db[IDXDB_TABLES.messages].get(message.parent);
+
+		if (!parent) return;
+
+		parent.children = parent.children.filter((childId: string) => childId !== messageId);
+		await db[IDXDB_TABLES.messages].put(parent);
+	}
+
+	/**
+	 * Reparents direct children of `parentId` to the nearest surviving
+	 * ancestor (or promotes them to top-level when the immediate parent was
+	 * top-level). Walking skips any ancestor listed in `excludeIds`, since
+	 * those will be deleted in the same batch — leaving a grandchild pointing
+	 * at an `excludeIds` entry would orphan it. Children whose own id is in
+	 * `excludeIds` are dropped from the updates (the bulk-delete pass will
+	 * remove them). `prefetched` may carry a pre-fetched ancestor map to
+	 * avoid repeat reads inside a bulk transaction.
+	 */
+	private static async reparentDirectChildren(
+		parentId: string,
+		excludeIds: ReadonlySet<string> = new Set(),
+		prefetched?: ReadonlyMap<string, DatabaseConversation>
+	): Promise<void> {
+		const conv = prefetched?.get(parentId) ?? (await db[IDXDB_TABLES.conversations].get(parentId));
+
+		if (!conv) return;
+
+		let newParent = conv.forkedFromConversationId;
+
+		const visited = new Set<string>([parentId]);
+
+		while (newParent && excludeIds.has(newParent)) {
+			if (visited.has(newParent)) {
+				newParent = undefined;
+
+				break;
+			}
+
+			visited.add(newParent);
+			const next =
+				prefetched?.get(newParent) ?? (await db[IDXDB_TABLES.conversations].get(newParent));
+
+			if (!next) {
+				newParent = undefined;
+
+				break;
+			}
+
+			newParent = next.forkedFromConversationId;
+		}
+
+		const directChildren = await db[IDXDB_TABLES.conversations]
+			.filter((c) => c.forkedFromConversationId === parentId)
+			.toArray();
+		const updates: DatabaseConversation[] = [];
+
+		for (const child of directChildren) {
+			if (excludeIds.has(child.id)) continue;
+
+			updates.push({ ...child, forkedFromConversationId: newParent });
+		}
+
+		if (updates.length === 0) return;
+
+		await db[IDXDB_TABLES.conversations].bulkPut(updates);
+	}
 }
@@ -53,9 +53,9 @@
 * - Reasoning content stripping from prompt history to avoid KV cache pollution
 * - Error translation (network, timeout, server errors → user-friendly messages)
 *
- * @see chatStore in stores/chat.svelte.ts — primary consumer for chat state management
- * @see agenticStore in stores/agentic.svelte.ts — uses ChatService for agentic loop streaming
- * @see conversationsStore in stores/conversations.svelte.ts — provides message context
+ * @see chatStore in stores/chat/index.svelte.ts — primary consumer for chat state management
+ * @see agenticStore in stores/agentic/index.svelte.ts — uses ChatService for agentic loop streaming
+ * @see conversationsStore in stores/conversations/index.svelte.ts — provides message context
 */
 export { ChatService } from './chat.service';

@@ -98,8 +98,8 @@ export { ChatService } from './chat.service';
 * enabling conversation branching and alternative response paths. The conversation's
 * `currNode` tracks the currently active branch endpoint.
 *
- * @see conversationsStore in stores/conversations.svelte.ts — reactive layer on top of DatabaseService
- * @see chatStore in stores/chat.svelte.ts — uses DatabaseService directly for message CRUD during streaming
+ * @see conversationsStore in stores/conversations/index.svelte.ts — reactive layer on top of DatabaseService
+ * @see chatStore in stores/chat/index.svelte.ts — uses DatabaseService directly for message CRUD during streaming
 */
 export { DatabaseService } from './database.service';

@@ -143,7 +143,7 @@ export { ConversationTransferService } from './conversation-transfer.service';
 * - `POST /models/load` — Load a model (ROUTER mode only)
 * - `POST /models/unload` — Unload a model (ROUTER mode only)
 *
- * @see modelsStore in stores/models.svelte.ts — primary consumer for reactive model state
+ * @see modelsStore in stores/models/index.svelte.ts — primary consumer for reactive model state
 */
 export { ModelsService } from './models.service';

@@ -174,8 +174,8 @@ export { ModelsService } from './models.service';
 * - `&autoload=false` → Prevents model auto-loading when querying props
 *
 * @see serverStore in stores/server.svelte.ts — consumes global server props
- * @see modelsStore in stores/models.svelte.ts — consumes per-model props for modalities
- * @see settingsStore in stores/settings.svelte.ts — syncs default generation params from props
+ * @see modelsStore in stores/models/index.svelte.ts — consumes per-model props for modalities
+ * @see settingsStore in stores/settings/index.svelte.ts — syncs default generation params from props
 */
 export { PropsService } from './props.service';

@@ -217,7 +217,7 @@ export { PropsService } from './props.service';
 * - `ParameterSyncService` class — static methods for sync logic
 * - `SYNCABLE_PARAMETERS` — mapping of UI setting keys to server parameter keys
 *
- * @see settingsStore in stores/settings.svelte.ts — primary consumer for settings sync
+ * @see settingsStore in stores/settings/index.svelte.ts — primary consumer for settings sync
 * @see SettingsChatParameterSourceIndicator — displays parameter source badges in UI
 */
 export { ParameterSyncService } from './parameter-sync.service';
@@ -241,7 +241,7 @@ export { ParameterSyncService } from './parameter-sync.service';
 *   - Manages connection lifecycle, health checks, reconnection
 *   - Handles tool name conflict resolution and server coordination
 *
- * - **mcpResourceStore**: Reactive resource state
+ * - **mcpResourceStore** (composed as mcpStore.resources): Reactive resource state
 *   - Receives resource data fetched via MCPService
 *   - Manages resource caching, subscriptions, and attachments
 *
@@ -263,9 +263,9 @@ export { ParameterSyncService } from './parameter-sync.service';
 * 2. **StreamableHTTP** — modern HTTP-based, supports CORS proxy
 * 3. **SSE** — legacy fallback, supports CORS proxy
 *
- * @see mcpStore in stores/mcp.svelte.ts — reactive business logic facade on top of MCPService
- * @see mcpResourceStore in stores/mcp-resources.svelte.ts — reactive resource state management
- * @see agenticStore in stores/agentic.svelte.ts — uses MCPService (via mcpStore) for tool execution
+ * @see mcpStore in stores/mcp/index.svelte.ts — reactive business logic facade on top of MCPService
+ * @see mcpStore.resources in stores/mcp/resources.svelte.ts — reactive resource state management
+ * @see agenticStore in stores/agentic/index.svelte.ts — uses MCPService (via mcpStore) for tool execution
 * @see MCP Protocol Specification: https://modelcontextprotocol.io/specification/2025-06-18
 */
 export { MCPService } from './mcp.service';
@@ -286,7 +286,7 @@ export { MCPService } from './mcp.service';
 * - **agenticStore**: Dispatches ToolSource.BROWSER calls here
 *
 * @see buildSandboxToolDefinition in utils/sandbox-tool - tool schema sent to the LLM
- * @see agenticStore in stores/agentic.svelte.ts - tool dispatch
+ * @see agenticStore in stores/agentic/index.svelte.ts - tool dispatch
 */
 export { SandboxService } from './sandbox.service';

@@ -340,3 +340,13 @@ export { RouterService } from './router.service';
 * @see migration.service.ts — full implementation (non-destructive)
 */
 export { MigrationService } from './migration.service';
+
+/**
+ * **SettingsService** - localStorage persistence layer for settings
+ *
+ * Stateless read/write of the settings config and user-override keys. Business
+ * logic (default merging, mobile defaults, theme migration) stays in the store.
+ *
+ * @see settingsStore in stores/settings/index.svelte.ts - reactive state + business logic
+ */
+export { SettingsService } from './settings.service';
@@ -1,20 +1,11 @@
 /**
- * Migration Service - Unified data migration hook
+ * MigrationService - Unified data migration hook
 *
- * Centralizes all data migrations (localStorage, IndexedDB, legacy formats) into a single
- * initialization point. Each migration copies data to new format WITHOUT deleting the old.
- *
- * **Architecture:**
- * - Migrations are defined as objects with `id` and `run()` methods
- * - Migration state is tracked in localStorage to avoid re-running
- * - `runAllMigrations()` should be called once at app startup
- * - All migrations are NON-DESTRUCTIVE - legacy data is preserved for downgrade compatibility
- *
- * **Current Migrations:**
- * 1. localStorage prefix: Copy LlamaCppWebui.* → LlamaUi.* (both preserved)
- * 2. IndexedDB database: Copy LlamacppWebui → LlamaUi (both preserved)
- * 3. Legacy message format: Transform in-place (preserves structure, migrates markers)
- * 4. Theme key: Copy standalone `theme` → config object (both preserved)
+ * Centralizes all data migrations (localStorage, IndexedDB, legacy formats)
+ * into a single initialization point. Each migration copies data to the new
+ * format WITHOUT deleting the old, and state is tracked in localStorage so
+ * `runAllMigrations()` (called once at startup) never re-runs a completed
+ * migration. All migrations are non-destructive for downgrade compatibility.
 */

 import {
@@ -1,25 +1,55 @@
+/**
+ * ModelsService - Stateless model management API layer
+ *
+ * Wraps the /models endpoints (list, load, unload) and the /models/sse
+ * status feed in MODEL and ROUTER modes. No reactive state; consumed by
+ * modelsStore and its status manager.
+ */
+
 import { base } from '$app/paths';
-import {
-	API_MODELS,
-	MODEL_ID,
-	SSE_DATA_PREFIX,
-	SSE_LINE_SEPARATOR,
-	SSE_RECORD_SEPARATOR
-} from '$lib/constants';
+import { API_MODELS, MODEL_ID } from '$lib/constants';
 import { ServerModelStatus } from '$lib/enums';
 import type { ParsedModelId } from '$lib/types/models';
-import { apiFetch, apiPost, normalizeModelName } from '$lib/utils';
+import {
+	apiFetch,
+	apiPost,
+	extractSseDataPayload,
+	normalizeModelName,
+	splitSseRecords
+} from '$lib/utils';
 import { getAuthHeaders } from '$lib/utils/api-headers';

 export class ModelsService {
+	private static readonly SSE_RECONNECT_MS = 1000;
+
+	/**
+	 * Check if a model is loaded based on its metadata.
+	 *
+	 * @param model - Model data entry from the API response
+	 * @returns True if the model status is LOADED
+	 */
+	static isModelLoaded(model: ApiModelDataEntry): boolean {
+		return model.status.value === ServerModelStatus.LOADED;
+	}
+
 	/**
 	 *
 	 *
-	 * Listing
+	 * Load/Unload
 	 *
 	 *
 	 */

+	/**
+	 * Check if a model is currently loading.
+	 *
+	 * @param model - Model data entry from the API response
+	 * @returns True if the model status is LOADING
+	 */
+	static isModelLoading(model: ApiModelDataEntry): boolean {
+		return model.status.value === ServerModelStatus.LOADING;
+	}
+
 	/**
 	 * Fetch list of models from OpenAI-compatible endpoint.
 	 * Works in both MODEL and ROUTER modes.
@@ -41,14 +71,6 @@ export class ModelsService {
 		return apiFetch<ApiRouterModelsListResponse>(API_MODELS.LIST);
 	}

-	/**
-	 *
-	 *
-	 * Load/Unload
-	 *
-	 *
-	 */
-
 	/**
 	 * Load a model (ROUTER mode only).
 	 * Sends POST request to `/models/load`. Note: the endpoint returns success
@@ -68,137 +90,6 @@ export class ModelsService {
 		return apiPost<ApiRouterModelsLoadResponse>(API_MODELS.LOAD, payload);
 	}

-	/**
-	 * Unload a model (ROUTER mode only).
-	 * Sends POST request to `/models/unload`. Note: the endpoint returns success
-	 * before unloading completes — use polling to await actual unload status.
-	 *
-	 * @param modelId - Model identifier to unload
-	 * @returns Unload response from the server
-	 */
-	static async unload(modelId: string): Promise<ApiRouterModelsUnloadResponse> {
-		return apiPost<ApiRouterModelsUnloadResponse>(API_MODELS.UNLOAD, { model: modelId });
-	}
-
-	/**
-	 *
-	 *
-	 * Status
-	 *
-	 *
-	 */
-
-	/**
-	 * Check if a model is loaded based on its metadata.
-	 *
-	 * @param model - Model data entry from the API response
-	 * @returns True if the model status is LOADED
-	 */
-	static isModelLoaded(model: ApiModelDataEntry): boolean {
-		return model.status.value === ServerModelStatus.LOADED;
-	}
-
-	/**
-	 * Check if a model is currently loading.
-	 *
-	 * @param model - Model data entry from the API response
-	 * @returns True if the model status is LOADING
-	 */
-	static isModelLoading(model: ApiModelDataEntry): boolean {
-		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;
-		}
-	}
-
-	/**
-	 *
-	 *
-	 * Parsing
-	 *
-	 *
-	 */
-
 	/**
 	 * Parse a model ID string into its structured components.
 	 *
@@ -311,4 +202,84 @@ export class ModelsService {

 		return result;
 	}
+
+	/**
+	 * Unload a model (ROUTER mode only).
+	 * Sends POST request to `/models/unload`. Note: the endpoint returns success
+	 * before unloading completes — use polling to await actual unload status.
+	 *
+	 * @param modelId - Model identifier to unload
+	 * @returns Unload response from the server
+	 */
+	static async unload(modelId: string): Promise<ApiRouterModelsUnloadResponse> {
+		return apiPost<ApiRouterModelsUnloadResponse>(API_MODELS.UNLOAD, { model: modelId });
+	}
+
+	/**
+	 * 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 });
+
+						const { records, rest } = splitSseRecords(buffer);
+
+						buffer = rest;
+
+						for (const record of records) {
+							const event = ModelsService.parseStatusRecord(record);
+
+							if (event) onEvent(event);
+						}
+					}
+				}
+			} 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 = extractSseDataPayload(record);
+
+		if (payload.length === 0) return null;
+
+		try {
+			return JSON.parse(payload) as ApiModelsSseEvent;
+		} catch {
+			return null;
+		}
+	}
 }
@@ -1,26 +1,71 @@
-import { SETTINGS_KEYS, SYNCABLE_PARAMETERS } from '$lib/constants';
+/**
+ * ParameterSyncService - Syncs sampling parameters with the server
+ *
+ * Decides for each sampling parameter whether the user's setting is an
+ * override of the server default, and normalizes floating-point values.
+ * No reactive state; consumed by settingsStore.
+ */
+
+import { SETTINGS_KEYS, SETTINGS_REGISTRY } from '$lib/constants';
 import { ParameterSource, SyncableParameterType } from '$lib/enums';
-import type { ParameterInfo, ParameterRecord, ParameterValue } from '$lib/types';
+import type { ParameterInfo, ParameterRecord, ParameterValue, SyncableParameter } from '$lib/types';
 import { normalizeFloatingPoint } from '$lib/utils';

+/** Mapping of UI setting keys to server parameter keys, derived from the registry. */
+export const SYNCABLE_PARAMETERS: SyncableParameter[] = SETTINGS_REGISTRY.flatMap(
+	(section) => section.settings
+)
+	.filter((s) => s.sync !== undefined)
+	.map((s) => ({
+		canSync: true,
+		key: s.key,
+		serverKey: s.sync!.serverKey,
+		type: s.sync!.paramType
+	}));
+
 export class ParameterSyncService {
 	/**
+	 * Check if a parameter can be synced from server.
 	 *
-	 *
-	 * Extraction
-	 *
-	 *
+	 * @param key - The parameter key to check
+	 * @returns True if the parameter is in the syncable parameters list
 	 */
+	static canSyncParameter(key: string): boolean {
+		return SYNCABLE_PARAMETERS.some((param) => param.key === key && param.canSync);
+	}

 	/**
-	 * Round floating-point numbers to avoid JavaScript precision issues.
-	 * E.g., 0.1 + 0.2 = 0.30000000000000004 → 0.3
+	 * Create a diff between current settings and server defaults.
+	 * Shows which parameters differ from server values, useful for debugging
+	 * and for the "Reset to defaults" functionality.
 	 *
-	 * @param value - Parameter value to normalize
-	 * @returns Precision-normalized value
+	 * @param currentSettings - Current parameter values in the settings store
+	 * @param serverDefaults - Default values extracted from server props
+	 * @returns Record of parameter diffs with current value, server value, and whether they differ
 	 */
-	private static roundFloatingPoint(value: ParameterValue): ParameterValue {
-		return normalizeFloatingPoint(value) as ParameterValue;
+	static createParameterDiff(
+		currentSettings: ParameterRecord,
+		serverDefaults: ParameterRecord
+	): Record<string, { current: ParameterValue; server: ParameterValue; differs: boolean }> {
+		const diff: Record<
+			string,
+			{ current: ParameterValue; server: ParameterValue; differs: boolean }
+		> = {};
+
+		for (const key of this.getSyncableParameterKeys()) {
+			const currentValue = currentSettings[key];
+			const serverValue = serverDefaults[key];
+
+			if (serverValue !== undefined) {
+				diff[key] = {
+					current: currentValue,
+					differs: currentValue !== serverValue,
+					server: serverValue
+				};
+			}
+		}
+
+		return diff;
 	}

 	/**
@@ -59,49 +104,6 @@ export class ParameterSyncService {
 		return extracted;
 	}

-	/**
-	 *
-	 *
-	 * Merging
-	 *
-	 *
-	 */
-
-	/**
-	 * Merge server defaults with current user settings.
-	 * User overrides always take priority — only parameters not in `userOverrides`
-	 * set will be updated from server defaults.
-	 *
-	 * @param currentSettings - Current parameter values in the settings store
-	 * @param serverDefaults - Default values extracted from server props
-	 * @param userOverrides - Set of parameter keys explicitly overridden by the user
-	 * @returns Merged parameter record with user overrides preserved
-	 */
-	static mergeWithServerDefaults(
-		currentSettings: ParameterRecord,
-		serverDefaults: ParameterRecord,
-		userOverrides: Set<string> = new Set()
-	): ParameterRecord {
-		const merged = { ...currentSettings };
-
-		for (const [key, serverValue] of Object.entries(serverDefaults)) {
-			// Only update if user hasn't explicitly overridden this parameter
-			if (!userOverrides.has(key)) {
-				merged[key] = this.roundFloatingPoint(serverValue);
-			}
-		}
-
-		return merged;
-	}
-
-	/**
-	 *
-	 *
-	 * Info
-	 *
-	 *
-	 */
-
 	/**
 	 * Get parameter information including source and values.
 	 * Used by SettingsChatParameterSourceIndicator to display the correct badge
@@ -132,16 +134,6 @@ export class ParameterSyncService {
 		};
 	}

-	/**
-	 * Check if a parameter can be synced from server.
-	 *
-	 * @param key - The parameter key to check
-	 * @returns True if the parameter is in the syncable parameters list
-	 */
-	static canSyncParameter(key: string): boolean {
-		return SYNCABLE_PARAMETERS.some((param) => param.key === key && param.canSync);
-	}
-
 	/**
 	 * Get all syncable parameter keys.
 	 *
@@ -151,6 +143,33 @@ export class ParameterSyncService {
 		return SYNCABLE_PARAMETERS.filter((param) => param.canSync).map((param) => param.key);
 	}

+	/**
+	 * Merge server defaults with current user settings.
+	 * User overrides always take priority — only parameters not in `userOverrides`
+	 * set will be updated from server defaults.
+	 *
+	 * @param currentSettings - Current parameter values in the settings store
+	 * @param serverDefaults - Default values extracted from server props
+	 * @param userOverrides - Set of parameter keys explicitly overridden by the user
+	 * @returns Merged parameter record with user overrides preserved
+	 */
+	static mergeWithServerDefaults(
+		currentSettings: ParameterRecord,
+		serverDefaults: ParameterRecord,
+		userOverrides: Set<string> = new Set()
+	): ParameterRecord {
+		const merged = { ...currentSettings };
+
+		for (const [key, serverValue] of Object.entries(serverDefaults)) {
+			// Only update if user hasn't explicitly overridden this parameter
+			if (!userOverrides.has(key)) {
+				merged[key] = this.roundFloatingPoint(serverValue);
+			}
+		}
+
+		return merged;
+	}
+
 	/**
 	 * Validate a server parameter value against its expected type.
 	 *
@@ -176,44 +195,13 @@ export class ParameterSyncService {
 	}

 	/**
+	 * Round floating-point numbers to avoid JavaScript precision issues.
+	 * E.g., 0.1 + 0.2 = 0.30000000000000004 → 0.3
 	 *
-	 *
-	 * Diff
-	 *
-	 *
+	 * @param value - Parameter value to normalize
+	 * @returns Precision-normalized value
 	 */
-
-	/**
-	 * Create a diff between current settings and server defaults.
-	 * Shows which parameters differ from server values, useful for debugging
-	 * and for the "Reset to defaults" functionality.
-	 *
-	 * @param currentSettings - Current parameter values in the settings store
-	 * @param serverDefaults - Default values extracted from server props
-	 * @returns Record of parameter diffs with current value, server value, and whether they differ
-	 */
-	static createParameterDiff(
-		currentSettings: ParameterRecord,
-		serverDefaults: ParameterRecord
-	): Record<string, { current: ParameterValue; server: ParameterValue; differs: boolean }> {
-		const diff: Record<
-			string,
-			{ current: ParameterValue; server: ParameterValue; differs: boolean }
-		> = {};
-
-		for (const key of this.getSyncableParameterKeys()) {
-			const currentValue = currentSettings[key];
-			const serverValue = serverDefaults[key];
-
-			if (serverValue !== undefined) {
-				diff[key] = {
-					current: currentValue,
-					differs: currentValue !== serverValue,
-					server: serverValue
-				};
-			}
-		}
-
-		return diff;
+	private static roundFloatingPoint(value: ParameterValue): ParameterValue {
+		return normalizeFloatingPoint(value) as ParameterValue;
 	}
 }
@@ -1,14 +1,14 @@
+/**
+ * PropsService - Fetches server properties from /props
+ *
+ * Returns global server settings and capabilities, including per-model
+ * modalities in MODEL mode. No reactive state; consumed by serverStore and
+ * the model props manager.
+ */
+
 import { apiFetchWithParams } from '$lib/utils';

 export class PropsService {
-	/**
-	 *
-	 *
-	 * Fetching
-	 *
-	 *
-	 */
-
 	/**
 	 * Fetches global server properties from the `/props` endpoint.
 	 * In MODEL mode, returns modalities for the single loaded model.
@@ -1,3 +1,10 @@
+/**
+ * ReadMediaService - Reads local media files for the read_media tool
+ *
+ * Encodes image and audio files as base64 data URLs with the metadata the
+ * model needs. No reactive state; consumed by toolsStore.
+ */
+
 import { ToolsService } from './tools.service';
 import {
 	FILE_EXTENSION_SEPARATOR,
@@ -40,7 +47,7 @@ function fileExtension(path: string): string {
 * actually use the result - the server has no idea which model is selected.
 *
 * @see buildReadMediaToolDefinition in constants/read-media.ts - tool schema sent to the LLM
- * @see agenticStore in stores/agentic.svelte.ts - tool dispatch and attachment extraction
+ * @see agenticStore in stores/agentic/index.svelte.ts - tool dispatch and attachment extraction
 */
 export class ReadMediaService {
 	static async executeTool(
@@ -1,3 +1,10 @@
+/**
+ * RouterService - Builds app route paths
+ *
+ * Returns chat and settings route strings from a single source of truth
+ * (ROUTES). No state.
+ */
+
 import { ROUTES } from '$lib/constants';

 export class RouterService {
@@ -1,3 +1,10 @@
+/**
+ * Sandbox harness - builds the srcdoc document for the sandboxed iframe
+ *
+ * Produces the HTML/CSP/worker shim that runs untrusted model code in an
+ * opaque origin. Consumed by sandbox.service.
+ */
+
 import WORKER_SHIM from './sandbox-worker.js?raw';
 import { NEWLINE } from '$lib/constants';

@@ -1,3 +1,11 @@
+/**
+ * SandboxService - Runs untrusted code in a sandboxed worker
+ *
+ * Executes model-generated code inside a CSP-restricted, opaque-origin
+ * iframe worker with output and timeout limits. No reactive state; consumed
+ * by toolsStore for code-execution tools.
+ */
+
 import { buildSandboxHarness } from './sandbox-harness';
 import {
 	NEWLINE,
@@ -8,7 +16,7 @@ import {
 	SANDBOX_TOOL_NAME,
 	SANDBOX_TRUNCATION_NOTICE
 } from '$lib/constants';
-import { settingsStore } from '$lib/stores/settings.svelte';
+import { settingsStore } from '$lib/stores/settings/index.svelte';
 import type { ToolExecutionResult } from '$lib/types';

 /** Cached harnesses keyed by whether nerdamer is included. */
@@ -0,0 +1,76 @@
+import { browser } from '$app/environment';
+import { CONFIG_LOCALSTORAGE_KEY, USER_OVERRIDES_LOCALSTORAGE_KEY } from '$lib/constants';
+
+/**
+ * SettingsService - localStorage persistence layer for settings
+ *
+ * Stateless read/write of the settings config and user-override keys. Business
+ * logic (default merging, mobile defaults, theme migration) stays in the store.
+ *
+ * **Architecture & Relationships:**
+ * - **settingsStore**: Primary consumer - loads config on init and persists on change
+ *
+ * @see settingsStore in stores/settings/index.svelte.ts - reactive state + business logic
+ */
+export class SettingsService {
+	/**
+	 * Read the raw config and user overrides from localStorage.
+	 * @returns Parsed values, or empty defaults when nothing is stored or parsing fails.
+	 */
+	static loadConfig(): {
+		config: Record<string, unknown>;
+		userOverrides: string[];
+		isFirstVisit: boolean;
+	} {
+		if (!browser) {
+			return { config: {}, isFirstVisit: false, userOverrides: [] };
+		}
+
+		try {
+			const storedConfigRaw = localStorage.getItem(CONFIG_LOCALSTORAGE_KEY);
+			const isFirstVisit = storedConfigRaw === null;
+			const config = JSON.parse(storedConfigRaw || '{}') as Record<string, unknown>;
+			const userOverrides = JSON.parse(
+				localStorage.getItem(USER_OVERRIDES_LOCALSTORAGE_KEY) || '[]'
+			) as string[];
+
+			return { config, isFirstVisit, userOverrides };
+		} catch (error) {
+			console.warn('Failed to parse config from localStorage, using defaults:', error);
+
+			return { config: {}, isFirstVisit: false, userOverrides: [] };
+		}
+	}
+
+	/**
+	 * Migrate the legacy un-namespaced "theme" localStorage key.
+	 * Returns the legacy theme value (and removes the key) when present, else null.
+	 */
+	static migrateLegacyTheme(): string | null {
+		if (!browser) return null;
+
+		const legacyTheme = localStorage.getItem('theme');
+
+		if (legacyTheme) {
+			localStorage.removeItem('theme');
+
+			return legacyTheme;
+		}
+
+		return null;
+	}
+
+	/**
+	 * Persist the config and user overrides to localStorage.
+	 */
+	static saveConfig(config: Record<string, unknown>, userOverrides: string[]): void {
+		if (!browser) return;
+
+		try {
+			localStorage.setItem(CONFIG_LOCALSTORAGE_KEY, JSON.stringify(config));
+			localStorage.setItem(USER_OVERRIDES_LOCALSTORAGE_KEY, JSON.stringify(userOverrides));
+		} catch (error) {
+			console.error('Failed to save config to localStorage:', error);
+		}
+	}
+}
--- a/Show More
+++ b/Show More