fix launch model picker recovery

mlx: configurable model load timeout (#14796 )
launch: rename Codex App integration to ChatGPT (#17161 )
2026-07-13 17:15:07 -07:00 · 2026-07-13 16:06:29 -07:00 · 2026-07-13 14:34:17 -07:00 · 2026-07-13 12:52:40 -07:00 · 2026-07-13 12:50:38 -07:00 · 2026-07-09 17:27:31 -07:00
1577 changed files with 133282 additions and 450430 deletions
@@ -16,7 +16,7 @@ jobs:
    outputs:
      GOFLAGS: ${{ steps.goflags.outputs.GOFLAGS }}
      VERSION: ${{ steps.goflags.outputs.VERSION }}
-      vendorsha: ${{ steps.changes.outputs.vendorsha }}
+      vendorsha: ${{ steps.goflags.outputs.vendorsha }}
    steps:
      - uses: actions/checkout@v4
      - name: Set environment
@@ -24,10 +24,10 @@ jobs:
        run: |
          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${GITHUB_REF_NAME#v}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" | tee -a $GITHUB_OUTPUT
          echo VERSION="${GITHUB_REF_NAME#v}" | tee -a $GITHUB_OUTPUT
-          echo vendorsha=$(make -f Makefile.sync print-base) | tee -a $GITHUB_OUTPUT
+          echo vendorsha=$(cat LLAMA_CPP_VERSION)-$(cat MLX_VERSION)-$(cat MLX_C_VERSION) | tee -a $GITHUB_OUTPUT

  darwin-build:
-    runs-on: macos-14-xlarge
+    runs-on: macos-26-xlarge
    environment: release
    needs: setup-environment
    env:
@@ -39,11 +39,27 @@ jobs:
      APPLE_ID: ${{ vars.APPLE_ID }}
      MACOS_SIGNING_KEY: ${{ secrets.MACOS_SIGNING_KEY }}
      MACOS_SIGNING_KEY_PASSWORD: ${{ secrets.MACOS_SIGNING_KEY_PASSWORD }}
+      DEVELOPER_DIR: /Applications/Xcode_26.4.1.app/Contents/Developer
      CGO_CFLAGS: '-mmacosx-version-min=14.0 -O3'
      CGO_CXXFLAGS: '-mmacosx-version-min=14.0 -O3'
      CGO_LDFLAGS: '-mmacosx-version-min=14.0 -O3'
    steps:
      - uses: actions/checkout@v4
+      - name: Select Xcode 26.4.1
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [ ! -d "${DEVELOPER_DIR}" ]; then
+            echo "Missing ${DEVELOPER_DIR}"
+            ls -1 /Applications | grep '^Xcode' || true
+            exit 1
+          fi
+
+          sudo xcode-select -s "${DEVELOPER_DIR}"
+          sw_vers
+          xcodebuild -version
+          xcrun --sdk macosx --show-sdk-version
+          xcrun --find metal
      - run: |
          echo $MACOS_SIGNING_KEY | base64 --decode > certificate.p12
          security create-keychain -p password build.keychain
@@ -57,7 +73,9 @@ jobs:
          go-version-file: go.mod
          cache-dependency-path: |
            go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
      - run: |
          ./scripts/build_darwin.sh
      - name: Log build results
@@ -73,15 +91,18 @@ jobs:
            dist/*.dmg

  windows-depends:
+    needs: setup-environment
    strategy:
      matrix:
        os: [windows]
        arch: [amd64]
        preset: ['CPU']
+        build-steps: ['cpu cpuArm64']
        include:
          - os: windows
            arch: amd64
            preset: 'CUDA 12'
+            build-steps: cuda12
            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
            cuda-components:
              - '"cudart"'
@@ -89,10 +110,10 @@ jobs:
              - '"cublas"'
              - '"cublas_dev"'
            cuda-version: '12.8'
-            flags: ''
          - os: windows
            arch: amd64
            preset: 'CUDA 13'
+            build-steps: cuda13
            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
            cuda-components:
              - '"cudart"'
@@ -103,23 +124,23 @@ jobs:
              - '"nvvm"'
              - '"nvptxcompiler"'
            cuda-version: '13.0'
-            flags: ''
          - os: windows
            arch: amd64
-            preset: 'ROCm 6'
-            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
-            rocm-version: '6.2'
-            flags: '-DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
-            runner_dir: 'rocm'
+            preset: 'ROCm 7'
+            build-steps: rocm7
+            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-26.Q1-Win11-For-HIP.exe
+            rocm-version: '7.1'
          - os: windows
            arch: amd64
            preset: Vulkan
+            build-steps: vulkan
            install: https://sdk.lunarg.com/sdk/download/1.4.321.1/windows/vulkansdk-windows-X64-1.4.321.1.exe
-            flags: ''
-            runner_dir: 'vulkan'
          - os: windows
            arch: amd64
            preset: 'MLX CUDA 13'
+            build-steps: mlxCuda13
+            build-parallel: '16'
+            cmake-cuda-flags: '-t 6'
            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
            cudnn-install: https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/windows-x86_64/cudnn-windows-x86_64-9.18.1.3_cuda13-archive.zip
            cuda-components:
@@ -135,18 +156,34 @@ jobs:
              - '"nvvm"'
              - '"nvptxcompiler"'
            cuda-version: '13.0'
-            flags: ''
    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
    environment: release
    env:
      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
+      VERSION: ${{ needs.setup-environment.outputs.VERSION }}
    steps:
+      - if: startsWith(matrix.preset, 'MLX ')
+        name: Increase pagefile to 200 GB
+        uses: al-cheb/configure-pagefile-action@v1.5
+        with:
+          minimum-size: 16GB
+          maximum-size: 200GB
+          disk-root: "D:"
      - name: Install system dependencies
        run: |
          choco install -y --no-progress ccache ninja
          if (Get-Command ccache -ErrorAction SilentlyContinue) {
            ccache -o cache_dir=${{ github.workspace }}\.ccache
          }
+      - if: matrix.preset == 'CPU'
+        name: Install Windows ARM64 cross compiler
+        run: |
+          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-x86_64.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt.zip"
+          Expand-Archive -Path ${{ runner.temp }}\llvm-mingw-ucrt.zip -DestinationPath "C:\Program Files\"
+          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt-x86_64").path
+          if (!(Test-Path "$installPath\bin\aarch64-w64-mingw32-gcc.exe")) {
+            throw "llvm-mingw x86_64 package is missing the aarch64 cross compiler"
+          }
      - if: startsWith(matrix.preset, 'CUDA ') || startsWith(matrix.preset, 'ROCm ') || startsWith(matrix.preset, 'Vulkan') || startsWith(matrix.preset, 'MLX ')
        id: cache-install
        uses: actions/cache/restore@v4
@@ -195,12 +232,12 @@ jobs:
          }
          
          $vulkanPath = (Resolve-Path "C:\VulkanSDK\*").path
+          $vulkanRuntime = Join-Path $vulkanPath "Helpers\VulkanRT.exe"
+          if (Test-Path $vulkanRuntime) {
+            Start-Process -FilePath $vulkanRuntime -ArgumentList "/s" -NoNewWindow -Wait
+          }
          echo "$vulkanPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
          echo "VULKAN_SDK=$vulkanPath" >> $env:GITHUB_ENV
-      - if: matrix.preset == 'CPU'
-        run: |
-          echo "CC=clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "CXX=clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
      - if: startsWith(matrix.preset, 'MLX ')
        name: Install cuDNN for MLX
        run: |
@@ -232,72 +269,63 @@ jobs:
        with:
          path: ${{ github.workspace }}\.ccache
          key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}-${{ needs.setup-environment.outputs.vendorsha }}
-      - name: Build target "${{ matrix.preset }}"
+      - name: Build Windows dependencies
        run: |
          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }} --install-prefix "$((pwd).Path)\dist\${{ matrix.os }}-${{ matrix.arch }}"
-          cmake --build --parallel ([Environment]::ProcessorCount) --preset "${{ matrix.preset }}"
-          cmake --install build --component "${{ startsWith(matrix.preset, 'MLX ') && 'MLX' || startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || startsWith(matrix.preset, 'Vulkan') && 'Vulkan' || 'CPU' }}" --strip
-          Remove-Item -Path dist\lib\ollama\rocm\rocblas\library\*gfx906* -ErrorAction SilentlyContinue
+          $steps = "${{ matrix.build-steps }}".Split(' ', [System.StringSplitOptions]::RemoveEmptyEntries)
+          ./scripts/build_windows.ps1 @steps
        env:
          CMAKE_GENERATOR: Ninja
+          OLLAMA_BUILD_PARALLEL: ${{ matrix.build-parallel || '' }}
+          OLLAMA_CMAKE_CUDA_FLAGS: ${{ matrix.cmake-cuda-flags || '' }}
      - name: Log build results
        run: |
          gci -path .\dist -Recurse -File | ForEach-Object { get-filehash -path $_.FullName -Algorithm SHA256 } | format-list
+      - if: matrix.preset == 'CPU'
+        name: Verify Windows CPU payloads
+        shell: bash
+        run: |
+          set -euo pipefail
+          for payload in \
+            dist/windows-amd64/lib/ollama/llama-server.exe \
+            dist/windows-arm64/lib/ollama/llama-server.exe
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
      - uses: actions/upload-artifact@v4
        with:
          name: depends-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}
          path: dist\*

  windows-build:
-    strategy:
-      matrix:
-        os: [windows]
-        arch: [amd64, arm64]
-        include:
-        - os: windows
-          arch: amd64
-          llvmarch: x86_64
-        - os: windows
-          arch: arm64
-          llvmarch: aarch64
-    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
+    runs-on: windows
    environment: release
    needs: [setup-environment]
    env:
      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
      VERSION: ${{ needs.setup-environment.outputs.VERSION }}
    steps:
-      - name: Install ARM64 system dependencies
-        if: matrix.arch == 'arm64'
-        run: |
-          $ErrorActionPreference = "Stop"
-          Set-ExecutionPolicy Bypass -Scope Process -Force
-          [System.Net.ServicePointManager]::SecurityProtocol = [System.Net.ServicePointManager]::SecurityProtocol -bor 3072
-          iex ((New-Object System.Net.WebClient).DownloadString('https://community.chocolatey.org/install.ps1'))
-          echo "C:\ProgramData\chocolatey\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-
-          Invoke-WebRequest -Uri https://aka.ms/vs/17/release/vc_redist.arm64.exe -OutFile "${{ runner.temp }}\vc_redist.arm64.exe"
-          Start-Process -FilePath "${{ runner.temp }}\vc_redist.arm64.exe" -ArgumentList @("/install", "/quiet", "/norestart") -NoNewWindow -Wait
-
-          choco install -y --no-progress git gzip
-          echo "C:\Program Files\Git\cmd" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
      - name: Install clang and gcc-compat
        run: |
          $ErrorActionPreference = "Stop"
          Set-ExecutionPolicy Bypass -Scope Process -Force
-          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-${{ matrix.llvmarch }}.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt.zip"
+          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-x86_64.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt.zip"
          Expand-Archive -Path ${{ runner.temp }}\llvm-mingw-ucrt.zip -DestinationPath "C:\Program Files\"
-          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt*").path
+          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt-x86_64").path
          echo "$installPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          if (!(Test-Path "$installPath\bin\aarch64-w64-mingw32-gcc.exe")) {
+            throw "llvm-mingw x86_64 package is missing the aarch64 cross compiler"
+          }
      - uses: actions/checkout@v4
      - uses: actions/setup-go@v5
        with:
          go-version-file: go.mod
          cache-dependency-path: |
            go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
      - name: Verify gcc is actually clang
        run: |
          $ErrorActionPreference='Continue'
@@ -314,20 +342,30 @@ jobs:
        with:
          node-version: "20"
      - run: |
-          ./scripts/build_windows ollama app
+          ./scripts/build_windows ollama ollamaArm64 app appArm64
+      - name: Verify Windows build payloads
+        shell: bash
+        run: |
+          set -euo pipefail
+          for payload in \
+            dist/windows-amd64/ollama.exe \
+            dist/windows-arm64/ollama.exe
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
      - name: Log build results
        run: |
          gci -path .\dist -Recurse -File | ForEach-Object { get-filehash -path $_.FullName -Algorithm SHA256 } | format-list
      - uses: actions/upload-artifact@v4
        with:
-          name: build-${{ matrix.os }}-${{ matrix.arch }}
+          name: build-windows-amd64
          path: |
            dist\*

  windows-app:
    runs-on: windows
    environment: release
-    needs: [windows-build, windows-depends]
+    needs: [setup-environment, windows-build, windows-depends]
    env:
      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
      VERSION: ${{ needs.setup-environment.outputs.VERSION }}
@@ -353,7 +391,9 @@ jobs:
          go-version-file: go.mod
          cache-dependency-path: |
            go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
      - uses: actions/download-artifact@v4
        with:
          pattern: depends-windows*
@@ -367,6 +407,18 @@ jobs:
      - name: Log dist contents after download
        run: |
          gci -path .\dist -recurse
+      - name: Verify Windows package inputs
+        shell: bash
+        run: |
+          set -euo pipefail
+          for payload in \
+            dist/windows-amd64/ollama.exe \
+            dist/windows-amd64/lib/ollama/llama-server.exe \
+            dist/windows-arm64/ollama.exe \
+            dist/windows-arm64/lib/ollama/llama-server.exe
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
      - run: |
          ./scripts/build_windows.ps1 deps sign installer zip
      - name: Log contents after build
@@ -380,20 +432,33 @@ jobs:
            dist/*.ps1
            dist/OllamaSetup.exe

-  linux-build:
+  linux-depends:
    strategy:
      matrix:
        include:
-          - os: linux
-            arch: amd64
-            target: archive
-          - os: linux
-            arch: amd64
-            target: rocm
-          - os: linux
-            arch: arm64
-            target: archive
-    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
+          - arch: amd64
+            target: llama-server-cpu
+          - arch: amd64
+            target: llama-server-cuda_v12
+          - arch: amd64
+            target: llama-server-cuda_v13
+          - arch: amd64
+            target: mlx
+          - arch: amd64
+            target: llama-server-rocm_v7_2
+          - arch: amd64
+            target: llama-server-vulkan
+          - arch: arm64
+            target: llama-server-cpu
+          - arch: arm64
+            target: llama-server-cuda_v12
+          - arch: arm64
+            target: llama-server-cuda_v13
+          - arch: arm64
+            target: jetpack-5
+          - arch: arm64
+            target: jetpack-6
+    runs-on: ${{ matrix.arch == 'arm64' && 'linux-arm64' || 'linux' }}
    environment: release
    needs: setup-environment
    env:
@@ -401,80 +466,114 @@ jobs:
    steps:
      - uses: actions/checkout@v4
      - uses: docker/setup-buildx-action@v3
+      - uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - if: matrix.target == 'mlx'
+        name: Increase Linux swap to 200 GB
+        shell: bash
+        run: |
+          set -e
+          SWAP_PATH=/swapfile-mlx
+          SWAP_SIZE_GB=200
+          if [ -f "$SWAP_PATH" ]; then
+              sudo swapoff "$SWAP_PATH" 2>/dev/null || true
+              sudo rm -f "$SWAP_PATH"
+          fi
+          if ! sudo fallocate -l ${SWAP_SIZE_GB}G "$SWAP_PATH" 2>/dev/null; then
+              echo "fallocate unsupported, falling back to dd"
+              sudo dd if=/dev/zero of="$SWAP_PATH" bs=1M count=$((SWAP_SIZE_GB * 1024))
+          fi
+          sudo chmod 600 "$SWAP_PATH"
+          sudo mkswap "$SWAP_PATH"
+          sudo swapon "$SWAP_PATH"
+          swapon --show
+          free -h
      - uses: docker/build-push-action@v6
        with:
          context: .
-          platforms: ${{ matrix.os }}/${{ matrix.arch }}
+          platforms: linux/${{ matrix.arch }}
          target: ${{ matrix.target }}
+          provenance: false
+          sbom: false
          build-args: |
            GOFLAGS=${{ env.GOFLAGS }}
            CGO_CFLAGS=${{ env.CGO_CFLAGS }}
            CGO_CXXFLAGS=${{ env.CGO_CXXFLAGS }}
-          outputs: type=local,dest=dist/${{ matrix.os }}-${{ matrix.arch }}
-          cache-from: type=registry,ref=${{ vars.DOCKER_REPO }}:latest
-          cache-to: type=inline
-      - name: Deduplicate CUDA libraries
-        run: |
-          ./scripts/deduplicate_cuda_libs.sh dist/${{ matrix.os }}-${{ matrix.arch }}
-      - run: |
-          for COMPONENT in bin/* lib/ollama/*; do
-            case "$COMPONENT" in
-              bin/ollama*)               echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/vulkan*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/mlx*)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/include*)       echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
-              lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
-              lib/ollama/rocm)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
-            esac
-          done
-        working-directory: dist/${{ matrix.os }}-${{ matrix.arch }}
-      - run: |
-          echo "Manifests"
-          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in ; do
-            echo $ARCHIVE
-            cat $ARCHIVE
-          done
-      - run: |
-          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
-            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | zstd --ultra -22 -T0 >$(basename ${ARCHIVE//.*/}.tar.zst);
-          done
-      - uses: actions/upload-artifact@v4
-        with:
-          name: bundles-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.target }}
-          path: |
-            *.tar.zst
+            OLLAMA_MLX_BUILD_JOBS=16
+            OLLAMA_MLX_NVCC_THREADS=6
+            APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+            APT_PORTS_MIRROR=http://azure.ports.ubuntu.com/ubuntu-ports
+          cache-from: |
+            type=registry,ref=ollama/release:cache-${{ matrix.arch }}-${{ matrix.target }}
+            type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+          cache-to: type=registry,ref=ollama/release:cache-${{ matrix.arch }}-${{ matrix.target }},mode=max

  # Build each Docker variant (OS, arch, and flavor) separately. Using QEMU is unreliable and slower.
+  # Heavy stages were pre-built by linux-depends; this job is cache-hit-only for those layers
+  # and just assembles, runs the Go build, pushes the final image, and extracts release bundles.
  docker-build-push:
    strategy:
      matrix:
        include:
          - os: linux
            arch: arm64
+            archive-target: archive
            build-args: |
              CGO_CFLAGS
              CGO_CXXFLAGS
              GOFLAGS
+              APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+              APT_PORTS_MIRROR=http://azure.ports.ubuntu.com/ubuntu-ports
+              OLLAMA_MLX_BUILD_JOBS=16
+              OLLAMA_MLX_NVCC_THREADS=6
+            cache-from: |
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cuda_v12
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cuda_v13
+              type=registry,ref=ollama/release:cache-arm64-jetpack-5
+              type=registry,ref=ollama/release:cache-arm64-jetpack-6
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
          - os: linux
            arch: amd64
+            archive-target: archive
            build-args: |
              CGO_CFLAGS
              CGO_CXXFLAGS
              GOFLAGS
+              APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+              APT_PORTS_MIRROR=http://azure.ports.ubuntu.com/ubuntu-ports
+              OLLAMA_MLX_BUILD_JOBS=16
+              OLLAMA_MLX_NVCC_THREADS=6
+            cache-from: |
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cuda_v12
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cuda_v13
+              type=registry,ref=ollama/release:cache-amd64-mlx
+              type=registry,ref=ollama/release:cache-amd64-llama-server-rocm_v7_2
+              type=registry,ref=ollama/release:cache-amd64-llama-server-vulkan
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
          - os: linux
            arch: amd64
            suffix: '-rocm'
+            archive-target: image-archive
            build-args: |
              CGO_CFLAGS
              CGO_CXXFLAGS
              GOFLAGS
              FLAVOR=rocm
+              APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+              APT_PORTS_MIRROR=http://azure.ports.ubuntu.com/ubuntu-ports
+              OLLAMA_MLX_BUILD_JOBS=16
+              OLLAMA_MLX_NVCC_THREADS=6
+            cache-from: |
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-amd64-llama-server-rocm_v7_2
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
    environment: release
-    needs: setup-environment
+    needs: [setup-environment, linux-depends]
    env:
      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
    steps:
@@ -489,9 +588,11 @@ jobs:
        with:
          context: .
          platforms: ${{ matrix.os }}/${{ matrix.arch }}
+          provenance: false
+          sbom: false
          build-args: ${{ matrix.build-args }}
          outputs: type=image,name=${{ vars.DOCKER_REPO }},push-by-digest=true,name-canonical=true,push=true
-          cache-from: type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+          cache-from: ${{ matrix.cache-from }}
          cache-to: type=inline
      - run: |
          mkdir -p ${{ matrix.os }}-${{ matrix.arch }}
@@ -502,6 +603,64 @@ jobs:
          name: digest-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.suffix }}
          path: |
            ${{ runner.temp }}/${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.suffix }}.txt
+      - uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: ${{ matrix.os }}/${{ matrix.arch }}
+          target: ${{ matrix.archive-target }}
+          provenance: false
+          sbom: false
+          build-args: ${{ matrix.build-args }}
+          outputs: type=local,dest=dist/${{ matrix.os }}-${{ matrix.arch }}
+          cache-from: ${{ matrix.cache-from }}
+      - name: Deduplicate CUDA libraries
+        run: |
+          ./scripts/deduplicate_cuda_libs.sh dist/${{ matrix.os }}-${{ matrix.arch }}
+      - name: Verify Linux build payloads
+        shell: bash
+        run: |
+          set -euo pipefail
+          base="dist/${{ matrix.os }}-${{ matrix.arch }}"
+          for payload in \
+            "$base/bin/ollama" \
+            "$base/lib/ollama/llama-server"
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
+      - run: |
+          for COMPONENT in bin/* lib/ollama/*; do
+            case "$COMPONENT" in
+              bin/ollama*)               echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/llama-server*|lib/ollama/llama-quantize*)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/vulkan*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/mlx*)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-mlx.tar.in ;;
+              lib/ollama/include*)       echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-mlx.tar.in ;;
+              lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
+              lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
+              lib/ollama/rocm_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
+            esac
+          done
+        working-directory: dist/${{ matrix.os }}-${{ matrix.arch }}
+      - if: matrix.suffix == '-rocm'
+        run: rm -f dist/${{ matrix.os }}-${{ matrix.arch }}/ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in
+      - run: |
+          echo "Manifests"
+          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in ; do
+            echo $ARCHIVE
+            cat $ARCHIVE
+          done
+      - run: |
+          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
+            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | zstd -19 -T0 >$(basename ${ARCHIVE//.*/}.tar.zst) &
+          done
+          wait
+      - uses: actions/upload-artifact@v4
+        with:
+          name: bundles-${{ matrix.os }}-${{ matrix.arch }}${{ matrix.suffix }}
+          path: |
+            *.tar.zst

  # Merge Docker images for the same flavor into a single multi-arch manifest
  docker-merge-push:
@@ -541,7 +700,7 @@ jobs:
  release:
    runs-on: ubuntu-latest
    environment: release
-    needs: [darwin-build, windows-app, linux-build]
+    needs: [darwin-build, windows-app, docker-build-push]
    permissions:
      contents: write
    env:
@@ -559,6 +718,21 @@ jobs:
      - name: Copy install scripts to dist
        run: |
          cp scripts/install.sh dist/install.sh
+      - name: Verify release artifacts
+        run: |
+          required=(
+            dist/OllamaSetup.exe
+            dist/install.ps1
+            dist/install.sh
+            dist/ollama-windows-amd64.zip
+            dist/ollama-windows-arm64.zip
+          )
+          for payload in "${required[@]}"; do
+            if [ ! -f "$payload" ]; then
+              echo "::error::Missing expected release artifact: $payload"
+              exit 1
+            fi
+          done
      - name: Generate checksum file
        run: find . -type f -not -name 'sha256sum.txt' | xargs sha256sum | tee sha256sum.txt
        working-directory: dist
@@ -0,0 +1,596 @@
+name: test-llamacpp-update
+
+# PR validation artifacts from this workflow are intentionally unsigned and not
+# notarized. They are for llama.cpp update testing only and must not be
+# published as release artifacts.
+
+on:
+  pull_request:
+    paths:
+      - 'LLAMA_CPP_VERSION'
+
+permissions:
+  contents: read
+
+env:
+  CGO_CFLAGS: '-O3'
+  CGO_CXXFLAGS: '-O3'
+
+jobs:
+  setup-environment:
+    runs-on: ubuntu-latest
+    outputs:
+      GOFLAGS: ${{ steps.goflags.outputs.GOFLAGS }}
+      VERSION: ${{ steps.goflags.outputs.VERSION }}
+      vendorsha: ${{ steps.goflags.outputs.vendorsha }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set environment
+        id: goflags
+        shell: bash
+        run: |
+          set -euo pipefail
+          VERSION="0.0.0-llamacpp-${GITHUB_SHA::7}"
+          {
+            echo "GOFLAGS='-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${VERSION}\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
+            echo "VERSION=${VERSION}"
+            echo "vendorsha=$(cat LLAMA_CPP_VERSION)-$(cat MLX_VERSION)-$(cat MLX_C_VERSION)"
+          } >>"${GITHUB_OUTPUT}"
+
+  darwin-build:
+    runs-on: macos-26-xlarge
+    needs: setup-environment
+    env:
+      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
+      VERSION: ${{ needs.setup-environment.outputs.VERSION }}
+      CGO_CFLAGS: '-mmacosx-version-min=14.0 -O3'
+      CGO_CXXFLAGS: '-mmacosx-version-min=14.0 -O3'
+      CGO_LDFLAGS: '-mmacosx-version-min=14.0 -O3'
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache-dependency-path: |
+            go.sum
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
+      - name: Build unsigned Darwin runtime
+        run: ./scripts/build_darwin.sh build package
+      - name: Log build results
+        run: ls -l dist/
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ollama-darwin.tgz
+          path: dist/ollama-darwin.tgz
+          compression-level: 0
+
+  # Build payload export stages independently and combine the exported
+  # filesystem artifacts below. This preserves parallelism without Docker
+  # registry credentials or oversized GitHub layer caches.
+  linux-payloads:
+    runs-on: ${{ matrix.arch == 'arm64' && 'linux-arm64' || 'linux' }}
+    needs: setup-environment
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - arch: amd64
+            target: publish-llama-server-cpu
+            payload: cpu
+          - arch: amd64
+            target: publish-llama-server-cuda_v12
+            payload: cuda_v12
+          - arch: amd64
+            target: publish-llama-server-cuda_v13
+            payload: cuda_v13
+          - arch: amd64
+            target: publish-llama-server-rocm_v7_2
+            payload: rocm_v7_2
+          - arch: amd64
+            target: publish-llama-server-vulkan
+            payload: vulkan
+          - arch: arm64
+            target: publish-llama-server-cpu
+            payload: cpu
+          - arch: arm64
+            target: publish-llama-server-cuda_v12
+            payload: cuda_v12
+          - arch: arm64
+            target: publish-llama-server-cuda_v13
+            payload: cuda_v13
+          - arch: arm64
+            target: publish-llama-server-cuda_jetpack5
+            payload: cuda_jetpack5
+          - arch: arm64
+            target: publish-llama-server-cuda_jetpack6
+            payload: cuda_jetpack6
+    steps:
+      - uses: actions/checkout@v4
+      - uses: docker/setup-buildx-action@v3
+      - uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: linux/${{ matrix.arch }}
+          target: ${{ matrix.target }}
+          provenance: false
+          sbom: false
+          build-args: |
+            GOFLAGS=${{ needs.setup-environment.outputs.GOFLAGS }}
+            CGO_CFLAGS=${{ env.CGO_CFLAGS }}
+            CGO_CXXFLAGS=${{ env.CGO_CXXFLAGS }}
+            APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+            APT_PORTS_MIRROR=http://azure.ports.ubuntu.com/ubuntu-ports
+          outputs: type=local,dest=${{ runner.temp }}/payload
+      - name: Pack Linux payload
+        shell: bash
+        run: |
+          set -euo pipefail
+          tar -C "${{ runner.temp }}/payload" -cf - . | zstd -9 -T0 >"${{ runner.temp }}/linux-payload-${{ matrix.arch }}-${{ matrix.payload }}.tar.zst"
+      - uses: actions/upload-artifact@v4
+        with:
+          name: linux-payload-${{ matrix.arch }}-${{ matrix.payload }}
+          path: ${{ runner.temp }}/linux-payload-${{ matrix.arch }}-${{ matrix.payload }}.tar.zst
+          compression-level: 0
+
+  linux-go:
+    runs-on: ${{ matrix.arch == 'arm64' && 'linux-arm64' || 'linux' }}
+    needs: setup-environment
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: [amd64, arm64]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: docker/setup-buildx-action@v3
+      - uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: linux/${{ matrix.arch }}
+          target: publish-go
+          provenance: false
+          sbom: false
+          build-args: |
+            GOFLAGS=${{ needs.setup-environment.outputs.GOFLAGS }}
+            CGO_CFLAGS=${{ env.CGO_CFLAGS }}
+            CGO_CXXFLAGS=${{ env.CGO_CXXFLAGS }}
+            APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+            APT_PORTS_MIRROR=http://azure.ports.ubuntu.com/ubuntu-ports
+          outputs: type=local,dest=${{ runner.temp }}/payload
+      - name: Pack Linux Go payload
+        shell: bash
+        run: |
+          set -euo pipefail
+          tar -C "${{ runner.temp }}/payload" -cf - . | zstd -9 -T0 >"${{ runner.temp }}/linux-payload-${{ matrix.arch }}-go.tar.zst"
+      - uses: actions/upload-artifact@v4
+        with:
+          name: linux-payload-${{ matrix.arch }}-go
+          path: ${{ runner.temp }}/linux-payload-${{ matrix.arch }}-go.tar.zst
+          compression-level: 0
+
+  # MLX payloads are intentionally excluded from this workflow; the Dockerfile
+  # still exposes publish-mlx for a separate MLX-specific workflow.
+  linux-bundles:
+    runs-on: ${{ matrix.arch == 'arm64' && 'linux-arm64' || 'linux' }}
+    needs: [linux-payloads, linux-go]
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: [amd64, arm64]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/download-artifact@v4
+        with:
+          pattern: linux-payload-${{ matrix.arch }}-*
+          path: ${{ runner.temp }}/payloads
+          merge-multiple: true
+      - name: Assemble Linux payload tree
+        shell: bash
+        run: |
+          set -euo pipefail
+          src="${{ runner.temp }}/payloads"
+          arch="${{ matrix.arch }}"
+          dest="dist/linux-${arch}"
+          copy_payload() {
+            local name="$1"
+            local payload="${src}/linux-payload-${arch}-${name}.tar.zst"
+            if [ ! -f "${payload}" ]; then
+              echo "missing payload ${payload}"
+              exit 1
+            fi
+            zstd -d <"${payload}" | tar -C "${dest}" -xf -
+          }
+
+          mkdir -p "${dest}"
+          copy_payload go
+          copy_payload cpu
+          copy_payload cuda_v12
+          copy_payload cuda_v13
+          if [ "${arch}" = "amd64" ]; then
+            copy_payload vulkan
+            copy_payload rocm_v7_2
+          else
+            copy_payload cuda_jetpack5
+            copy_payload cuda_jetpack6
+          fi
+
+          ./scripts/deduplicate_cuda_libs.sh "${dest}"
+      - name: Verify Linux build payloads
+        shell: bash
+        run: |
+          set -euo pipefail
+          base="dist/linux-${{ matrix.arch }}"
+          for payload in \
+            "${base}/bin/ollama" \
+            "${base}/lib/ollama/llama-server"
+          do
+            [ -f "${payload}" ] || { echo "missing ${payload}"; exit 1; }
+          done
+      - name: Create archive input lists
+        shell: bash
+        run: |
+          set -euo pipefail
+          for COMPONENT in bin/* lib/ollama/*; do
+            case "${COMPONENT}" in
+              bin/ollama*)              echo "${COMPONENT}" >>ollama-linux-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/*.so*)         echo "${COMPONENT}" >>ollama-linux-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/llama-server*|lib/ollama/llama-quantize*) echo "${COMPONENT}" >>ollama-linux-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_v*)       echo "${COMPONENT}" >>ollama-linux-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/vulkan*)       echo "${COMPONENT}" >>ollama-linux-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/mlx*)          echo "${COMPONENT}" >>ollama-linux-${{ matrix.arch }}-mlx.tar.in ;;
+              lib/ollama/include*)      echo "${COMPONENT}" >>ollama-linux-${{ matrix.arch }}-mlx.tar.in ;;
+              lib/ollama/cuda_jetpack5) echo "${COMPONENT}" >>ollama-linux-${{ matrix.arch }}-jetpack5.tar.in ;;
+              lib/ollama/cuda_jetpack6) echo "${COMPONENT}" >>ollama-linux-${{ matrix.arch }}-jetpack6.tar.in ;;
+              lib/ollama/rocm_v*)       echo "${COMPONENT}" >>ollama-linux-${{ matrix.arch }}-rocm.tar.in ;;
+            esac
+          done
+        working-directory: dist/linux-${{ matrix.arch }}
+      - name: Log archive input lists
+        shell: bash
+        run: |
+          set -euo pipefail
+          for ARCHIVE in dist/linux-${{ matrix.arch }}/*.tar.in; do
+            echo "${ARCHIVE}"
+            cat "${ARCHIVE}"
+          done
+      - name: Create Linux archives
+        shell: bash
+        run: |
+          set -euo pipefail
+          for ARCHIVE in dist/linux-${{ matrix.arch }}/*.tar.in; do
+            tar c -C dist/linux-${{ matrix.arch }} -T "${ARCHIVE}" --owner 0 --group 0 | zstd -19 -T0 >"$(basename "${ARCHIVE//.*/}.tar.zst")" &
+          done
+          wait
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ollama-linux-${{ matrix.arch }}.tar.zst
+          path: ollama-linux-${{ matrix.arch }}.tar.zst
+          compression-level: 0
+      - if: matrix.arch == 'amd64'
+        uses: actions/upload-artifact@v4
+        with:
+          name: ollama-linux-amd64-rocm.tar.zst
+          path: ollama-linux-amd64-rocm.tar.zst
+          compression-level: 0
+      - if: matrix.arch == 'arm64'
+        uses: actions/upload-artifact@v4
+        with:
+          name: ollama-linux-arm64-jetpack5.tar.zst
+          path: ollama-linux-arm64-jetpack5.tar.zst
+          compression-level: 0
+      - if: matrix.arch == 'arm64'
+        uses: actions/upload-artifact@v4
+        with:
+          name: ollama-linux-arm64-jetpack6.tar.zst
+          path: ollama-linux-arm64-jetpack6.tar.zst
+          compression-level: 0
+
+  windows-depends:
+    needs: setup-environment
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [windows]
+        arch: [amd64]
+        preset: ['CPU']
+        build-steps: ['cpu cpuArm64']
+        include:
+          - os: windows
+            arch: amd64
+            preset: 'CUDA 12'
+            build-steps: cuda12
+            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
+            cuda-version: '12.8'
+          - os: windows
+            arch: amd64
+            preset: 'CUDA 13'
+            build-steps: cuda13
+            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
+              - '"crt"'
+              - '"nvvm"'
+              - '"nvptxcompiler"'
+            cuda-version: '13.0'
+          - os: windows
+            arch: amd64
+            preset: 'ROCm 7'
+            build-steps: rocm7
+            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-26.Q1-Win11-For-HIP.exe
+            rocm-version: '7.1'
+          - os: windows
+            arch: amd64
+            preset: Vulkan
+            build-steps: vulkan
+            install: https://sdk.lunarg.com/sdk/download/1.4.321.1/windows/vulkansdk-windows-X64-1.4.321.1.exe
+    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
+    env:
+      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
+      VERSION: ${{ needs.setup-environment.outputs.VERSION }}
+    steps:
+      - name: Install system dependencies
+        run: |
+          choco install -y --no-progress ccache ninja
+          if (Get-Command ccache -ErrorAction SilentlyContinue) {
+            ccache -o cache_dir=${{ github.workspace }}\.ccache
+          }
+      - if: matrix.preset == 'CPU'
+        name: Install Windows ARM64 cross compiler
+        run: |
+          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-x86_64.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt.zip"
+          Expand-Archive -Path ${{ runner.temp }}\llvm-mingw-ucrt.zip -DestinationPath "C:\Program Files\"
+          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt-x86_64").path
+          if (!(Test-Path "$installPath\bin\aarch64-w64-mingw32-gcc.exe")) {
+            throw "llvm-mingw x86_64 package is missing the aarch64 cross compiler"
+          }
+      - if: startsWith(matrix.preset, 'CUDA ') || startsWith(matrix.preset, 'ROCm ') || startsWith(matrix.preset, 'Vulkan')
+        id: cache-install
+        uses: actions/cache/restore@v4
+        with:
+          path: |
+            C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
+            C:\Program Files\AMD\ROCm
+            C:\VulkanSDK
+          key: ${{ matrix.install }}
+      - if: startsWith(matrix.preset, 'CUDA ')
+        name: Install CUDA ${{ matrix.cuda-version }}
+        run: |
+          $ErrorActionPreference = "Stop"
+          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
+            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
+            $subpackages = @(${{ join(matrix.cuda-components, ', ') }}) | Foreach-Object {"${_}_${{ matrix.cuda-version }}"}
+            Start-Process -FilePath .\install.exe -ArgumentList (@("-s") + $subpackages) -NoNewWindow -Wait
+          }
+
+          $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path
+          echo "$cudaPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - if: startsWith(matrix.preset, 'ROCm')
+        name: Install ROCm ${{ matrix.rocm-version }}
+        run: |
+          $ErrorActionPreference = "Stop"
+          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
+            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
+            Start-Process -FilePath .\install.exe -ArgumentList '-install' -NoNewWindow -Wait
+          }
+
+          $hipPath = (Resolve-Path "C:\Program Files\AMD\ROCm\*").path
+          echo "$hipPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CC=$hipPath\bin\clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "CXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "HIPCXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "HIP_PLATFORM=amd" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "CMAKE_PREFIX_PATH=$hipPath" | Out-File -FilePath $env:GITHUB_ENV -Append
+      - if: matrix.preset == 'Vulkan'
+        name: Install Vulkan
+        run: |
+          $ErrorActionPreference = "Stop"
+          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
+            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
+            Start-Process -FilePath .\install.exe -ArgumentList "-c","--am","--al","in" -NoNewWindow -Wait
+          }
+
+          $vulkanPath = (Resolve-Path "C:\VulkanSDK\*").path
+          $vulkanRuntime = Join-Path $vulkanPath "Helpers\VulkanRT.exe"
+          if (Test-Path $vulkanRuntime) {
+            Start-Process -FilePath $vulkanRuntime -ArgumentList "/s" -NoNewWindow -Wait
+          }
+          echo "$vulkanPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "VULKAN_SDK=$vulkanPath" >> $env:GITHUB_ENV
+      - if: ${{ !cancelled() && matrix.preset != 'CPU' && steps.cache-install.outputs.cache-hit != 'true' }}
+        uses: actions/cache/save@v4
+        with:
+          path: |
+            C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
+            C:\Program Files\AMD\ROCm
+            C:\VulkanSDK
+          key: ${{ matrix.install }}
+      - uses: actions/checkout@v4
+      - uses: actions/cache@v4
+        with:
+          path: ${{ github.workspace }}\.ccache
+          key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}-${{ needs.setup-environment.outputs.vendorsha }}
+      - name: Build Windows dependencies
+        run: |
+          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation -DevCmdArguments '-arch=x64 -no_logo'
+          $steps = "${{ matrix.build-steps }}".Split(' ', [System.StringSplitOptions]::RemoveEmptyEntries)
+          ./scripts/build_windows.ps1 @steps
+        env:
+          CMAKE_GENERATOR: Ninja
+      - name: Log build results
+        run: |
+          gci -path .\dist -Recurse -File | ForEach-Object { get-filehash -path $_.FullName -Algorithm SHA256 } | format-list
+      - if: matrix.preset == 'CPU'
+        name: Verify Windows CPU payloads
+        shell: bash
+        run: |
+          set -euo pipefail
+          for payload in \
+            dist/windows-amd64/lib/ollama/llama-server.exe \
+            dist/windows-arm64/lib/ollama/llama-server.exe
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
+      - uses: actions/upload-artifact@v4
+        with:
+          name: depends-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}
+          path: dist\*
+          compression-level: 0
+
+  windows-build:
+    runs-on: windows
+    needs: setup-environment
+    env:
+      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
+      VERSION: ${{ needs.setup-environment.outputs.VERSION }}
+    steps:
+      - name: Install clang and gcc-compat
+        run: |
+          $ErrorActionPreference = "Stop"
+          Set-ExecutionPolicy Bypass -Scope Process -Force
+          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-x86_64.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt.zip"
+          Expand-Archive -Path ${{ runner.temp }}\llvm-mingw-ucrt.zip -DestinationPath "C:\Program Files\"
+          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt-x86_64").path
+          echo "$installPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          if (!(Test-Path "$installPath\bin\aarch64-w64-mingw32-gcc.exe")) {
+            throw "llvm-mingw x86_64 package is missing the aarch64 cross compiler"
+          }
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache-dependency-path: |
+            go.sum
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
+      - name: Verify gcc is actually clang
+        run: |
+          $ErrorActionPreference='Continue'
+          $version=& gcc -v 2>&1
+          $version=$version -join "`n"
+          echo "gcc is $version"
+          if ($version -notmatch 'clang') {
+            echo "ERROR: GCC must be clang for proper utf16 handling"
+            exit 1
+          }
+          $ErrorActionPreference='Stop'
+      - name: Setup Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: "20"
+      - name: Build Windows binaries and app launchers
+        run: ./scripts/build_windows.ps1 ollama ollamaArm64 app appArm64
+      - name: Verify Windows build payloads
+        shell: bash
+        run: |
+          set -euo pipefail
+          for payload in \
+            dist/windows-amd64/ollama.exe \
+            dist/windows-arm64/ollama.exe
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
+      - name: Log build results
+        run: |
+          gci -path .\dist -Recurse -File | ForEach-Object { get-filehash -path $_.FullName -Algorithm SHA256 } | format-list
+      - uses: actions/upload-artifact@v4
+        with:
+          name: build-windows-amd64
+          path: dist\*
+          compression-level: 0
+
+  windows-package:
+    runs-on: windows
+    needs: [setup-environment, windows-build, windows-depends]
+    env:
+      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
+      VERSION: ${{ needs.setup-environment.outputs.VERSION }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache-dependency-path: |
+            go.sum
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
+      - uses: actions/download-artifact@v4
+        with:
+          pattern: depends-windows*
+          path: dist
+          merge-multiple: true
+      - uses: actions/download-artifact@v4
+        with:
+          pattern: build-windows*
+          path: dist
+          merge-multiple: true
+      - name: Copy unsigned install script
+        run: Copy-Item -Path .\scripts\install.ps1 -Destination .\dist\install.ps1 -ErrorAction Stop
+      - name: Log dist contents after download
+        run: gci -path .\dist -recurse
+      - name: Verify Windows package inputs
+        shell: bash
+        run: |
+          set -euo pipefail
+          for payload in \
+            dist/windows-amd64/ollama.exe \
+            dist/windows-amd64/lib/ollama/llama-server.exe \
+            dist/windows-arm64/ollama.exe \
+            dist/windows-arm64/lib/ollama/llama-server.exe
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
+      - name: Build unsigned Windows installer and zips
+        run: ./scripts/build_windows.ps1 deps installer zip
+      - name: Log contents after build
+        run: |
+          gci -path .\dist -Recurse -File | ForEach-Object { get-filehash -path $_.FullName -Algorithm SHA256 } | format-list
+      - name: Verify Windows package outputs
+        shell: bash
+        run: |
+          set -euo pipefail
+          for payload in \
+            dist/ollama-windows-amd64.zip \
+            dist/ollama-windows-arm64.zip \
+            dist/ollama-windows-amd64-rocm.zip \
+            dist/OllamaSetup.exe \
+            dist/install.ps1
+          do
+            [ -f "$payload" ] || { echo "missing $payload"; exit 1; }
+          done
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ollama-windows-amd64.zip
+          path: dist/ollama-windows-amd64.zip
+          compression-level: 0
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ollama-windows-arm64.zip
+          path: dist/ollama-windows-arm64.zip
+          compression-level: 0
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ollama-windows-amd64-rocm.zip
+          path: dist/ollama-windows-amd64-rocm.zip
+          compression-level: 0
+      - uses: actions/upload-artifact@v4
+        with:
+          name: OllamaSetup.exe
+          path: dist/OllamaSetup.exe
+          compression-level: 0
+      - uses: actions/upload-artifact@v4
+        with:
+          name: install.ps1
+          path: dist/install.ps1
+          compression-level: 0
@@ -22,7 +22,8 @@ jobs:
    runs-on: ubuntu-latest
    outputs:
      changed: ${{ steps.changes.outputs.changed }}
-      vendorsha: ${{ steps.changes.outputs.vendorsha }}
+      app_changed: ${{ steps.changes.outputs.app_changed }}
+      enginehash: ${{ steps.changes.outputs.enginehash }}
    steps:
      - uses: actions/checkout@v4
        with:
@@ -37,8 +38,42 @@ jobs:
              | xargs python3 -c "import sys; from pathlib import Path; print(any(Path(x).match(glob) for x in sys.argv[1:] for glob in '$*'.split(' ')))"
          }

-          echo changed=$(changed 'llama/llama.cpp/**/*' 'ml/backend/ggml/ggml/**/*' '.github/**/*') | tee -a $GITHUB_OUTPUT
-          echo vendorsha=$(make -f Makefile.sync print-base) | tee -a $GITHUB_OUTPUT
+          echo changed=$(changed \
+            'CMakeLists.txt' \
+            'CMakePresets.json' \
+            'cmake/**' \
+            'cmake/**/*' \
+            'llama/server/**/*' \
+            'llama/compat/**/*' \
+            'LLAMA_CPP_VERSION' \
+            'MLX_VERSION' \
+            'MLX_C_VERSION' \
+            'llama/llama.cpp/**/*' \
+            'ml/backend/ggml/ggml/**/*' \
+            'x/imagegen/mlx/**' \
+            'x/imagegen/mlx/**/*' \
+            '.github/**/*') | tee -a $GITHUB_OUTPUT
+          echo app_changed=$(changed 'app/**' 'app/**/*') | tee -a $GITHUB_OUTPUT
+          echo enginehash=$(cat LLAMA_CPP_VERSION)-$(cat MLX_VERSION)-$(cat MLX_C_VERSION) | tee -a $GITHUB_OUTPUT
+
+  patches:
+    strategy:
+      matrix:
+        os: [ubuntu-latest, windows-latest]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Verify patches apply cleanly
+        shell: bash
+        run: |
+          cmake -S llama/server -B "$RUNNER_TEMP/llama-server-patch-check" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DBUILD_SHARED_LIBS=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_OPENMP=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DOLLAMA_RUNNER_DIR=

  linux:
    needs: [changes]
@@ -47,23 +82,41 @@ jobs:
      matrix:
        include:
          - preset: CPU
+            superbuild_target: ollama-local
+            superbuild_dir: build/local-superbuild
+            superbuild_args: ''
+            expected_payload: lib/ollama/llama-server
+            install-go: true
          - preset: CUDA
            container: nvidia/cuda:13.0.0-devel-ubuntu22.04
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
+            superbuild_target: ollama-llama-server-cuda_v13
+            superbuild_dir: build/local-superbuild-cuda_v13
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=cuda_v13 -DCMAKE_CUDA_ARCHITECTURES=87'
+            expected_payload: lib/ollama/cuda_v13/libggml-cuda.so
          - preset: ROCm
-            container: rocm/dev-ubuntu-22.04:7.2
+            container: rocm/dev-ubuntu-22.04:7.2.1
            extra-packages: rocm-libs
-            flags: '-DAMDGPU_TARGETS=gfx1010 -DCMAKE_PREFIX_PATH=/opt/rocm'
+            superbuild_target: ollama-llama-server-rocm_v7_2
+            superbuild_dir: build/local-superbuild-rocm_v7_2
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=rocm_v7_2 -DAMDGPU_TARGETS=gfx1010 -DCMAKE_PREFIX_PATH=/opt/rocm'
+            expected_payload: lib/ollama/rocm_v7_2/libggml-hip.so
          - preset: Vulkan
            container: ubuntu:22.04
            extra-packages: >
              mesa-vulkan-drivers vulkan-tools
              libvulkan1 libvulkan-dev
-              vulkan-sdk cmake ccache g++ make
+              vulkan-sdk spirv-headers cmake ccache g++ make
+            superbuild_target: ollama-llama-server-vulkan
+            superbuild_dir: build/local-superbuild-vulkan
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=vulkan'
+            expected_payload: lib/ollama/vulkan/libggml-vulkan.so
          - preset: 'MLX CUDA 13'
            container: nvidia/cuda:13.0.0-devel-ubuntu22.04
            extra-packages: libcudnn9-dev-cuda-13 libopenblas-dev liblapack-dev liblapacke-dev git curl
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=87 -DBLAS_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu -DLAPACK_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu'
+            superbuild_target: ollama-mlx-cuda_v13
+            superbuild_dir: build/local-superbuild-mlx-cuda_v13
+            superbuild_args: '-DOLLAMA_MLX_BACKENDS=cuda_v13 -DCMAKE_CUDA_ARCHITECTURES=87 -DMLX_CUDA_ARCHITECTURES=80-virtual -DBLAS_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu -DLAPACK_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu'
+            expected_payload: lib/ollama/mlx_cuda_v13/libmlx.so
            install-go: true
    runs-on: linux
    container: ${{ matrix.container }}
@@ -80,11 +133,9 @@ jobs:
            echo "deb [signed-by=/usr/share/keyrings/lunarg-archive-keyring.gpg]  https://packages.lunarg.com/vulkan/1.4.313 jammy main" | $sudo tee /etc/apt/sources.list.d/lunarg-vulkan-1.4.313-jammy.list > /dev/null
            $sudo apt-get update
          fi
-          $sudo apt-get install -y cmake ccache ${{ matrix.extra-packages }}
-          # MLX requires CMake 3.25+, install from official releases
-          if [ "${{ matrix.preset }}" = "MLX CUDA 13" ]; then
-            curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.31.2/cmake-3.31.2-linux-$(uname -m).tar.gz | $sudo tar xz -C /usr/local --strip-components 1
-          fi
+          $sudo apt-get install -y cmake ccache curl git ${{ matrix.extra-packages }}
+          # Use a current CMake for upstream llama.cpp and Vulkan dependency discovery.
+          curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.31.2/cmake-3.31.2-linux-$(uname -m).tar.gz | $sudo tar xz -C /usr/local --strip-components 1
          # Export VULKAN_SDK if provided by LunarG package (defensive)
          if [ -d "/usr/lib/x86_64-linux-gnu/vulkan" ] && [ "${{ matrix.preset }}" = "Vulkan" ]; then
            echo "VULKAN_SDK=/usr" >> $GITHUB_ENV
@@ -94,17 +145,30 @@ jobs:
      - if: matrix.install-go
        name: Install Go
        run: |
+          [ -n "${{ matrix.container }}" ] || sudo=sudo
          GO_VERSION=$(awk '/^go / { print $2 }' go.mod)
-          curl -fsSL "https://golang.org/dl/go${GO_VERSION}.linux-$(dpkg --print-architecture).tar.gz" | tar xz -C /usr/local
+          curl -fsSL "https://golang.org/dl/go${GO_VERSION}.linux-$(dpkg --print-architecture).tar.gz" | $sudo tar xz -C /usr/local
          echo "/usr/local/go/bin" >> $GITHUB_PATH
      - uses: actions/cache@v4
        with:
          path: /github/home/.cache/ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.vendorsha }}
-      - run: |
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
-          cmake --build --preset "${{ matrix.preset }}" --parallel
-
+          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.enginehash }}
+      - name: Build native superbuild
+        if: matrix.superbuild_target
+        run: |
+          cmake -S . -B "${{ matrix.superbuild_dir }}" ${{ matrix.superbuild_args }}
+          CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) cmake --build "${{ matrix.superbuild_dir }}" --target "${{ matrix.superbuild_target }}" -- -l $(nproc)
+          test -e "${{ matrix.superbuild_dir }}/${{ matrix.expected_payload }}"
+      - name: Verify local superbuild install
+        if: matrix.superbuild_target == 'ollama-local'
+        run: |
+          ./ollama --version
+          "${{ matrix.superbuild_dir }}/lib/ollama/llama-server" --version
+          test -x "${{ matrix.superbuild_dir }}/lib/ollama/llama-quantize"
+          cmake --install "${{ matrix.superbuild_dir }}" --component ollama-local --prefix "$RUNNER_TEMP/ollama-local"
+          "$RUNNER_TEMP/ollama-local/bin/ollama" --version
+          "$RUNNER_TEMP/ollama-local/lib/ollama/llama-server" --version
+          test -x "$RUNNER_TEMP/ollama-local/lib/ollama/llama-quantize"
  windows:
    needs: [changes]
    if: needs.changes.outputs.changed == 'True'
@@ -112,9 +176,16 @@ jobs:
      matrix:
        include:
          - preset: CPU
+            superbuild_target: ollama-local
+            superbuild_dir: build\local-superbuild
+            superbuild_args: ''
+            expected_payload: lib\ollama\llama-server.exe
          - preset: CUDA
            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
+            superbuild_target: ollama-llama-server-cuda_v13
+            superbuild_dir: build\local-superbuild-cuda_v13
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=cuda_v13 -DCMAKE_CUDA_ARCHITECTURES=80'
+            expected_payload: lib\ollama\cuda_v13\ggml-cuda.dll
            cuda-components:
              - '"cudart"'
              - '"nvcc"'
@@ -125,14 +196,26 @@ jobs:
              - '"nvptxcompiler"'
            cuda-version: '13.0'
          - preset: ROCm
-            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
-            flags: '-DAMDGPU_TARGETS=gfx1010 -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
+            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-26.Q1-Win11-For-HIP.exe
+            rocm-version: '7.1'
+            superbuild_target: ollama-llama-server-rocm_v7_1
+            superbuild_dir: build\local-superbuild-rocm_v7_1
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=rocm_v7_1 -DAMDGPU_TARGETS=gfx1010'
+            expected_payload: lib\ollama\rocm_v7_1\ggml-hip.dll
          - preset: Vulkan
            install: https://sdk.lunarg.com/sdk/download/1.4.321.1/windows/vulkansdk-windows-X64-1.4.321.1.exe
+            superbuild_target: ollama-llama-server-vulkan
+            superbuild_dir: build\local-superbuild-vulkan
+            superbuild_args: '-DOLLAMA_LLAMA_BACKENDS=vulkan'
+            expected_payload: lib\ollama\vulkan\ggml-vulkan.dll
          - preset: 'MLX CUDA 13'
            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
            cudnn-install: https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/windows-x86_64/cudnn-windows-x86_64-9.18.1.3_cuda13-archive.zip
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
+            superbuild_target: ollama-mlx-cuda_v13
+            superbuild_dir: build\local-superbuild-mlx-cuda_v13
+            superbuild_args: '-DOLLAMA_MLX_BACKENDS=cuda_v13 -DCMAKE_CUDA_ARCHITECTURES=80 -DMLX_CUDA_ARCHITECTURES=80-virtual'
+            expected_payload: lib\ollama\mlx_cuda_v13\mlx.dll
+            install-go: true
            cuda-components:
              - '"cudart"'
              - '"nvcc"'
@@ -201,6 +284,10 @@ jobs:
          }

          $vulkanPath = (Resolve-Path "C:\VulkanSDK\*").path
+          $vulkanRuntime = Join-Path $vulkanPath "Helpers\VulkanRT.exe"
+          if (Test-Path $vulkanRuntime) {
+            Start-Process -FilePath $vulkanRuntime -ArgumentList "/s" -NoNewWindow -Wait
+          }
          echo "$vulkanPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
          echo "VULKAN_SDK=$vulkanPath" >> $env:GITHUB_ENV
      - if: matrix.preset == 'MLX CUDA 13'
@@ -230,18 +317,44 @@ jobs:
            C:\Program Files\NVIDIA\CUDNN
          key: ${{ matrix.install }}-${{ matrix.cudnn-install }}
      - uses: actions/checkout@v4
+      - if: matrix.superbuild_target == 'ollama-local' || matrix.install-go
+        uses: actions/setup-go@v5
+        with:
+          go-version-file: 'go.mod'
      - uses: actions/cache@v4
        with:
          path: ${{ github.workspace }}\.ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.vendorsha }}
-      - run: |
+          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.enginehash }}
+      - name: Build native superbuild
+        if: matrix.superbuild_target
+        run: |
+          $ErrorActionPreference = "Stop"
          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
-          cmake --build --parallel --preset "${{ matrix.preset }}"
+          cmake -S . -B "${{ matrix.superbuild_dir }}" ${{ matrix.superbuild_args }}
+          $env:CMAKE_BUILD_PARALLEL_LEVEL = [Environment]::ProcessorCount
+          cmake --build "${{ matrix.superbuild_dir }}" --target "${{ matrix.superbuild_target }}" -- -l $([Environment]::ProcessorCount)
+          if (!(Test-Path "${{ matrix.superbuild_dir }}\${{ matrix.expected_payload }}")) {
+            throw "missing ${{ matrix.expected_payload }}"
+          }
        env:
          CMAKE_GENERATOR: Ninja
-
+      - name: Verify local superbuild install
+        if: matrix.superbuild_target == 'ollama-local'
+        run: |
+          $ErrorActionPreference = "Stop"
+          & ".\ollama.exe" --version
+          & "${{ matrix.superbuild_dir }}\lib\ollama\llama-server.exe" --version
+          if (!(Test-Path "${{ matrix.superbuild_dir }}\lib\ollama\llama-quantize.exe")) {
+            throw "missing llama-quantize.exe"
+          }
+          $installPrefix = Join-Path $env:RUNNER_TEMP "ollama-local"
+          cmake --install "${{ matrix.superbuild_dir }}" --component ollama-local --prefix "$installPrefix"
+          & "$installPrefix\bin\ollama.exe" --version
+          & "$installPrefix\lib\ollama\llama-server.exe" --version
+          if (!(Test-Path "$installPrefix\lib\ollama\llama-quantize.exe")) {
+            throw "missing installed llama-quantize.exe"
+          }
  go_mod_tidy:
    runs-on: ubuntu-latest
    steps:
@@ -250,6 +363,7 @@ jobs:
        run: go mod tidy --diff || (echo "Please run 'go mod tidy'." && exit 1)

  test:
+    needs: [changes]
    strategy:
      matrix:
        os: [ubuntu-latest, macos-latest, windows-latest]
@@ -263,7 +377,9 @@ jobs:
          go-version-file: 'go.mod'
          cache-dependency-path: |
            go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
+            MLX_VERSION
+            MLX_C_VERSION
      - uses: actions/setup-node@v4
        with:
          node-version: '20'
@@ -277,6 +393,17 @@ jobs:
        if: ${{ startsWith(matrix.os, 'ubuntu') }}
        working-directory: ./app/ui/app
        run: npm test
+      - name: Verify MLX generated files are current
+        if: ${{ startsWith(matrix.os, 'ubuntu') }}
+        run: |
+          cmake -S . -B build/mlx-generate -DOLLAMA_MLX_BACKENDS=cuda_v13
+          cmake --build build/mlx-generate --target ollama-mlx-generate-wrappers
+          git diff --exit-code -- \
+            x/imagegen/mlx/mlx.h \
+            x/imagegen/mlx/mlx.c \
+            x/mlxrunner/mlx/generated.h \
+            x/mlxrunner/mlx/generated.c \
+            x/mlxrunner/mlx/include/mlx/c
      - name: Run go generate
        run: go generate ./...

@@ -284,15 +411,10 @@ jobs:
        if: always()
        run: go test -count=1 -benchtime=1x ./...

+      - name: go test app with live updater tag
+        if: ${{ needs.changes.outputs.app_changed == 'True' && contains(fromJSON('["macos-latest","windows-latest"]'), matrix.os) }}
+        run: go test -count=1 -tags updater_live ./app/...
+
      - uses: golangci/golangci-lint-action@v9
        with:
          only-new-issues: true
-
-  patches:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-      - name: Verify patches apply cleanly and do not change files
-        run: |
-          make -f Makefile.sync clean checkout apply-patches sync
-          git diff --compact-summary --exit-code
@@ -15,3 +15,4 @@ __debug_bin*
 llama/build
 llama/vendor
 /ollama
+integration/testdata/models/
@@ -0,0 +1,21 @@
+# AGENTS.md
+
+## Building
+
+For a full build from the repository root:
+
+```sh
+cmake -B build .
+cmake --build build --parallel 8
+./ollama serve
+```
+
+For quick Go-only iteration against an existing native payload:
+
+```sh
+go build .
+go run . serve
+```
+
+See `docs/development.md` for prerequisites, platform notes, GPU backends, and
+the full development workflow.
@@ -0,0 +1,3 @@
+# CLAUDE.md
+
+See `AGENTS.md` for the shared agent instructions for this repository.
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.21)
+cmake_minimum_required(VERSION 3.24)

 project(Ollama C CXX)

@@ -23,30 +23,23 @@ include(GNUInstallDirs)

 find_package(Threads REQUIRED)

-set(CMAKE_BUILD_TYPE Release)
-set(BUILD_SHARED_LIBS ON)
+if(NOT CMAKE_CONFIGURATION_TYPES AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+endif()
+
+# These defaults can be overridden by presets (e.g., for static macOS llama-server builds)
+if(NOT DEFINED BUILD_SHARED_LIBS)
+    set(BUILD_SHARED_LIBS ON)
+endif()

 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX Requires gnu++17 extensions to compile properly
+set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX require gnu++17 extensions to compile properly

-set(GGML_BUILD ON)
-set(GGML_SHARED ON)
-set(GGML_CCACHE ON)
-set(GGML_BACKEND_DL ON)
-set(GGML_BACKEND_SHARED ON)
-set(GGML_SCHED_MAX_COPIES 4)
-
-set(GGML_LLAMAFILE ON)
-set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
-set(GGML_CUDA_GRAPHS ON)
-set(GGML_CUDA_FA ON)
-set(GGML_CUDA_COMPRESSION_MODE default)
-
-if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
-    OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))
-    set(GGML_CPU_ALL_VARIANTS ON)
-endif()
+# GGML backend for inference is provided by llama-server (built separately via
+# llama/server/CMakeLists.txt using FetchContent from the pinned llama.cpp source).
+# The root CMake project is the orchestration entrypoint; backend-specific
+# build rules live in subprojects under cmake/.

 if(APPLE)
    set(CMAKE_BUILD_RPATH "@loader_path")
@@ -55,7 +48,8 @@ if(APPLE)
 endif()

 set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
-set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama/${OLLAMA_RUNNER_DIR})
+set(OLLAMA_LIB_DIR "lib/ollama" CACHE STRING "Install destination for Ollama runtime payloads")
+set(OLLAMA_INSTALL_DIR ${OLLAMA_LIB_DIR}/${OLLAMA_RUNNER_DIR})

 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
@@ -64,301 +58,9 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_RELEASE ${OLLAMA_BUILD_DIR})

-# Store ggml include paths for use with target_include_directories later.
-# We avoid global include_directories() to prevent polluting the include path
-# for other projects like MLX (whose openblas dependency has its own common.h).
-set(GGML_INCLUDE_DIRS
-    ${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src
-    ${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/include
-    ${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu
-    ${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu/amx
-)
-
-add_compile_definitions(NDEBUG GGML_VERSION=0x0 GGML_COMMIT=0x0)
-
-# Define GGML version variables for shared library SOVERSION
-# These are required by ggml/src/CMakeLists.txt for proper library versioning
-set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 0)
-set(GGML_VERSION_PATCH 0)
-set(GGML_VERSION "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
-
-set(GGML_CPU ON)
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
-set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)
-
-get_target_property(CPU_VARIANTS ggml-cpu MANUALLY_ADDED_DEPENDENCIES)
-if(NOT CPU_VARIANTS)
-    set(CPU_VARIANTS "ggml-cpu")
-endif()
-
-# Apply ggml include directories to ggml targets only (not globally)
-target_include_directories(ggml-base PRIVATE ${GGML_INCLUDE_DIRS})
-foreach(variant ${CPU_VARIANTS})
-    if(TARGET ${variant})
-        target_include_directories(${variant} PRIVATE ${GGML_INCLUDE_DIRS})
-    endif()
-endforeach()
-
-install(TARGETS ggml-base ${CPU_VARIANTS}
-    RUNTIME_DEPENDENCIES
-        PRE_EXCLUDE_REGEXES ".*"
-    RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-    LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-    FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-)
-
-check_language(CUDA)
-if(CMAKE_CUDA_COMPILER)
-    if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.24" AND NOT CMAKE_CUDA_ARCHITECTURES)
-        set(CMAKE_CUDA_ARCHITECTURES "native")
-    endif()
-
-    find_package(CUDAToolkit)
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cuda)
-    target_include_directories(ggml-cuda PRIVATE ${GGML_INCLUDE_DIRS})
-    install(TARGETS ggml-cuda
-        RUNTIME_DEPENDENCIES
-            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
-            PRE_INCLUDE_REGEXES cublas cublasLt cudart
-            PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
-    )
-endif()
-
-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(908|90a|1200|1201):xnack[+-]$"
-    CACHE STRING
-    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(908|90a|1200|1201):xnack[+-]$\"."
-)
-
-check_language(HIP)
-if(CMAKE_HIP_COMPILER)
-    set(HIP_PLATFORM "amd")
-
-    if(NOT AMDGPU_TARGETS)
-        find_package(hip REQUIRED)
-        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(94[012]|101[02]|1030|110[012]|120[01])$")
-    endif()
-
-    if(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
-        list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
-    endif()
-
-    if(AMDGPU_TARGETS)
-        find_package(hip REQUIRED)
-        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)
-        target_include_directories(ggml-hip PRIVATE ${GGML_INCLUDE_DIRS})
-
-        if (WIN32)
-            target_compile_definitions(ggml-hip PRIVATE GGML_CUDA_NO_PEER_COPY)
-        endif()
-
-        target_compile_definitions(ggml-hip PRIVATE GGML_HIP_NO_VMM)
-
-        install(TARGETS ggml-hip
-            RUNTIME_DEPENDENCY_SET rocm
-            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-        )
-        install(RUNTIME_DEPENDENCY_SET rocm
-                DIRECTORIES ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR}
-                PRE_INCLUDE_REGEXES hipblas rocblas amdhip64 rocsolver amd_comgr hsa-runtime64 rocsparse tinfo rocprofiler-register roctx64 rocroller drm drm_amdgpu numa elf
-                PRE_EXCLUDE_REGEXES ".*"
-                POST_EXCLUDE_REGEXES "system32"
-            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-        )
-
-        foreach(HIP_LIB_BIN_INSTALL_DIR IN ITEMS ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR})
-            if(EXISTS ${HIP_LIB_BIN_INSTALL_DIR}/rocblas)
-                install(DIRECTORY ${HIP_LIB_BIN_INSTALL_DIR}/rocblas DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP)
-                break()
-            endif()
-        endforeach()
-    endif()
-endif()
-
-if(NOT APPLE)
-    find_package(Vulkan)
-    if(Vulkan_FOUND)
-        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
-        target_include_directories(ggml-vulkan PRIVATE ${GGML_INCLUDE_DIRS})
-        install(TARGETS ggml-vulkan
-            RUNTIME_DEPENDENCIES
-                PRE_INCLUDE_REGEXES vulkan
-                PRE_EXCLUDE_REGEXES ".*"
-            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-        )
-    endif()
-endif()
-
-option(MLX_ENGINE "Enable MLX backend" OFF)
-if(MLX_ENGINE)
-    message(STATUS "Setting up MLX (this takes a while...)")
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/imagegen/mlx)
-
-    # Find CUDA toolkit if MLX is built with CUDA support
-    find_package(CUDAToolkit)
-
-    # Build list of directories for runtime dependency resolution
-    set(MLX_RUNTIME_DIRS ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR})
-    # Add cuDNN bin paths for DLLs (Windows MLX CUDA builds)
-    # CUDNN_ROOT_DIR is the standard CMake variable for cuDNN location
-    if(DEFINED ENV{CUDNN_ROOT_DIR})
-        # cuDNN 9.x has versioned subdirectories under bin/ (e.g., bin/13.0/)
-        file(GLOB CUDNN_BIN_SUBDIRS "$ENV{CUDNN_ROOT_DIR}/bin/*")
-        list(APPEND MLX_RUNTIME_DIRS ${CUDNN_BIN_SUBDIRS})
-    endif()
-    # Add build output directory and MLX dependency build directories
-    list(APPEND MLX_RUNTIME_DIRS ${OLLAMA_BUILD_DIR})
-    # OpenBLAS DLL location (pre-built zip extracts into openblas-src/bin/)
-    list(APPEND MLX_RUNTIME_DIRS ${CMAKE_BINARY_DIR}/_deps/openblas-src/bin)
-    # NCCL: on Linux, if real NCCL is found, cmake bundles libnccl.so via the
-    # regex below. If NCCL is not found, MLX links a static stub (OBJECT lib)
-    # so there is no runtime dependency. This path covers the stub build dir
-    # for windows so we include the DLL in our dependencies.
-    list(APPEND MLX_RUNTIME_DIRS ${CMAKE_BINARY_DIR}/_deps/mlx-build/mlx/distributed/nccl/nccl_stub-prefix/src/nccl_stub-build/Release)
-
-    # Base regexes for runtime dependencies (cross-platform)
-    set(MLX_INCLUDE_REGEXES cublas cublasLt cudart cufft nvrtc nvrtc-builtins cudnn nccl openblas gfortran)
-    # On Windows, also include dl.dll (dlfcn-win32 POSIX emulation layer)
-    if(WIN32)
-        list(APPEND MLX_INCLUDE_REGEXES "^dl\\.dll$")
-    endif()
-
-    install(TARGETS mlx mlxc
-        RUNTIME_DEPENDENCIES
-            DIRECTORIES ${MLX_RUNTIME_DIRS}
-            PRE_INCLUDE_REGEXES ${MLX_INCLUDE_REGEXES}
-            PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
-        FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
-    )
-
-    # Install the Metal library for macOS arm64 (must be colocated with the binary)
-    # Metal backend is only built for arm64, not x86_64
-    if(APPLE AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
-        install(FILES ${CMAKE_BINARY_DIR}/_deps/mlx-build/mlx/backend/metal/kernels/mlx.metallib
-            DESTINATION ${OLLAMA_INSTALL_DIR}
-            COMPONENT MLX)
-    endif()
-
-    # Install headers for NVRTC JIT compilation at runtime.
-    # MLX's own install rules use the default component so they get skipped by
-    # --component MLX. Headers are installed alongside libmlx in OLLAMA_INSTALL_DIR.
-    #
-    # Layout:
-    #   ${OLLAMA_INSTALL_DIR}/include/cccl/{cuda,nv}/  — CCCL headers
-    #   ${OLLAMA_INSTALL_DIR}/include/*.h               — CUDA toolkit headers
-    #
-    # MLX's jit_module.cpp resolves CCCL via
-    #   current_binary_dir()[.parent_path()] / "include" / "cccl"
-    # On Linux, MLX's jit_module.cpp resolves CCCL via
-    #   current_binary_dir().parent_path() / "include" / "cccl", so we create a
-    #   symlink from lib/ollama/include -> ${OLLAMA_RUNNER_DIR}/include
-    #   This will need refinement if we add multiple CUDA versions for MLX in the future.
-    # CUDA runtime headers are found via CUDA_PATH env var (set by mlxrunner).
-    if(EXISTS ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/cuda)
-        install(DIRECTORY ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/cuda
-            DESTINATION ${OLLAMA_INSTALL_DIR}/include/cccl
-            COMPONENT MLX)
-        install(DIRECTORY ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/nv
-            DESTINATION ${OLLAMA_INSTALL_DIR}/include/cccl
-            COMPONENT MLX)
-        if(NOT WIN32 AND NOT APPLE)
-            install(CODE "
-                set(_link \"${CMAKE_INSTALL_PREFIX}/lib/ollama/include\")
-                set(_target \"${OLLAMA_RUNNER_DIR}/include\")
-                if(NOT EXISTS \${_link})
-                    execute_process(COMMAND \${CMAKE_COMMAND} -E create_symlink \${_target} \${_link})
-                endif()
-            " COMPONENT MLX)
-        endif()
-    endif()
-
-    # Install minimal CUDA toolkit headers needed by MLX JIT kernels.
-    # These are the transitive closure of includes from mlx/backend/cuda/device/*.cuh.
-    # The Go mlxrunner sets CUDA_PATH to OLLAMA_INSTALL_DIR so MLX finds them at
-    # $CUDA_PATH/include/*.h via NVRTC --include-path.
-    if(CUDAToolkit_FOUND)
-        # CUDAToolkit_INCLUDE_DIRS may be a semicolon-separated list
-        # (e.g. ".../include;.../include/cccl"). Find the entry that
-        # contains the CUDA runtime headers we need.
-        set(_cuda_inc "")
-        foreach(_dir ${CUDAToolkit_INCLUDE_DIRS})
-            if(EXISTS "${_dir}/cuda_runtime_api.h")
-                set(_cuda_inc "${_dir}")
-                break()
-            endif()
-        endforeach()
-        if(NOT _cuda_inc)
-            message(WARNING "Could not find cuda_runtime_api.h in CUDAToolkit_INCLUDE_DIRS: ${CUDAToolkit_INCLUDE_DIRS}")
-        else()
-            set(_dst "${OLLAMA_INSTALL_DIR}/include")
-            set(_MLX_JIT_CUDA_HEADERS
-                builtin_types.h
-                cooperative_groups.h
-                cuda_bf16.h
-                cuda_bf16.hpp
-                cuda_device_runtime_api.h
-                cuda_fp16.h
-                cuda_fp16.hpp
-                cuda_fp8.h
-                cuda_fp8.hpp
-                cuda_runtime_api.h
-                device_types.h
-                driver_types.h
-                math_constants.h
-                surface_types.h
-                texture_types.h
-                vector_functions.h
-                vector_functions.hpp
-                vector_types.h
-            )
-            foreach(_hdr ${_MLX_JIT_CUDA_HEADERS})
-                install(FILES "${_cuda_inc}/${_hdr}"
-                    DESTINATION ${_dst}
-                    COMPONENT MLX)
-            endforeach()
-            # Subdirectory headers
-            install(DIRECTORY "${_cuda_inc}/cooperative_groups"
-                DESTINATION ${_dst}
-                COMPONENT MLX
-                FILES_MATCHING PATTERN "*.h")
-            install(FILES "${_cuda_inc}/crt/host_defines.h"
-                DESTINATION "${_dst}/crt"
-                COMPONENT MLX)
-        endif()
-    endif()
-
-    # On Windows, explicitly install dl.dll (dlfcn-win32 POSIX dlopen emulation)
-    # RUNTIME_DEPENDENCIES auto-excludes it via POST_EXCLUDE_FILES_STRICT because
-    # dlfcn-win32 is a known CMake target with its own install rules (which install
-    # to the wrong destination). We must install it explicitly here.
-    if(WIN32)
-        install(FILES ${OLLAMA_BUILD_DIR}/dl.dll
-            DESTINATION ${OLLAMA_INSTALL_DIR}
-            COMPONENT MLX)
-    endif()
-
-    # Manually install CUDA runtime libraries that MLX loads via dlopen
-    # (not detected by RUNTIME_DEPENDENCIES since they aren't link-time deps)
-    if(CUDAToolkit_FOUND)
-        file(GLOB MLX_CUDA_LIBS
-            "${CUDAToolkit_LIBRARY_DIR}/libcudart.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libcublas.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libcublasLt.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libnvrtc.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libnvrtc-builtins.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libcufft.so*"
-            "${CUDAToolkit_LIBRARY_DIR}/libcudnn.so*")
-        if(MLX_CUDA_LIBS)
-            install(FILES ${MLX_CUDA_LIBS}
-                DESTINATION ${OLLAMA_INSTALL_DIR}
-                COMPONENT MLX)
-        endif()
-    endif()
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/llama/server/CMakeLists.txt")
+    set(OLLAMA_HAVE_LLAMA_SERVER TRUE)
+else()
+    set(OLLAMA_HAVE_LLAMA_SERVER FALSE)
 endif()
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/local.cmake)
@@ -11,109 +11,10 @@
      }
    },
    {
-      "name": "CPU",
-      "inherits": [ "Default" ]
-    },
-    {
-      "name": "CUDA",
-      "inherits": [ "Default" ]
-    },
-    {
-      "name": "CUDA 11",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50-virtual;60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual",
-        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2",
-        "OLLAMA_RUNNER_DIR": "cuda_v11"
-      }
-    },
-    {
-      "name": "CUDA 12",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50;52;60;61;70;75;80;86;89;90;90a;120",
-        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2",
-        "OLLAMA_RUNNER_DIR": "cuda_v12"
-      }
-    },
-    {
-      "name": "CUDA 13",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;103-virtual;110-virtual;120-virtual;121-virtual",
-        "CMAKE_CUDA_FLAGS": "-t 4",
-        "OLLAMA_RUNNER_DIR": "cuda_v13"
-      }
-    },
-    {
-      "name": "JetPack 5",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "72;87",
-        "OLLAMA_RUNNER_DIR": "cuda_jetpack5"
-      }
-    },
-    {
-      "name": "JetPack 6",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "87",
-        "OLLAMA_RUNNER_DIR": "cuda_jetpack6"
-      }
-    },
-    {
-      "name": "ROCm",
+      "name": "MLX Metal",
      "inherits": [ "Default" ],
      "cacheVariables": {
-        "CMAKE_HIP_PLATFORM": "amd"
-      }
-    },
-    {
-      "name": "ROCm 6",
-      "inherits": [ "ROCm" ],
-      "cacheVariables": {
-        "CMAKE_HIP_FLAGS": "-parallel-jobs=4",
-        "AMDGPU_TARGETS": "gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-",
-        "OLLAMA_RUNNER_DIR": "rocm"
-      }
-    },
-    {
-      "name": "ROCm 7",
-      "inherits": [ "ROCm" ],
-      "cacheVariables": {
-        "CMAKE_HIP_FLAGS": "-parallel-jobs=4",
-        "AMDGPU_TARGETS": "gfx942;gfx950;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1200;gfx1201;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-",
-        "OLLAMA_RUNNER_DIR": "rocm"
-      }
-    },
-    {
-      "name": "Vulkan",
-      "inherits": [ "Default" ],
-      "cacheVariables": {
-        "OLLAMA_RUNNER_DIR": "vulkan"
-      }
-    },
-    {
-      "name": "MLX",
-      "inherits": [ "Default" ],
-      "cacheVariables": {
-        "MLX_ENGINE": "ON",
-        "OLLAMA_RUNNER_DIR": "mlx"
-      }
-    },
-    {
-      "name": "MLX CUDA 12",
-      "inherits": [ "MLX", "CUDA 12" ],
-      "cacheVariables": {
-        "OLLAMA_RUNNER_DIR": "mlx_cuda_v12"
-      }
-    },
-    {
-      "name": "MLX CUDA 13",
-      "inherits": [ "MLX", "CUDA 13" ],
-      "cacheVariables": {
-        "MLX_CUDA_ARCHITECTURES": "86;89;90;90a;100;103;75-virtual;80-virtual;110-virtual;120-virtual;121-virtual",
-        "OLLAMA_RUNNER_DIR": "mlx_cuda_v13"
+        "OLLAMA_MLX_BACKENDS": "metal_v3;metal_v4"
      }
    }
  ],
@@ -124,74 +25,9 @@
      "configuration": "Release"
    },
    {
-      "name": "CPU",
-      "configurePreset": "Default",
-      "targets": [ "ggml-cpu" ]
-    },
-    {
-      "name": "CUDA",
-      "configurePreset": "CUDA",
-      "targets": [ "ggml-cuda" ]
-    },
-    {
-      "name": "CUDA 11",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "CUDA 11"
-    },
-    {
-      "name": "CUDA 12",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "CUDA 12"
-    },
-    {
-      "name": "CUDA 13",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "CUDA 13"
-    },
-    {
-      "name": "JetPack 5",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "JetPack 5"
-    },
-    {
-      "name": "JetPack 6",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "JetPack 6"
-    },
-    {
-      "name": "ROCm",
-      "configurePreset": "ROCm",
-      "targets": [ "ggml-hip" ]
-    },
-    {
-      "name": "ROCm 6",
-      "inherits": [ "ROCm" ],
-      "configurePreset": "ROCm 6"
-    },
-    {
-      "name": "ROCm 7",
-      "inherits": [ "ROCm" ],
-      "configurePreset": "ROCm 7"
-    },
-    {
-      "name": "Vulkan",
-      "targets": [ "ggml-vulkan" ],
-      "configurePreset": "Vulkan"
-    },
-    {
-      "name": "MLX",
-      "targets": [ "mlx", "mlxc" ],
-      "configurePreset": "MLX"
-    },
-    {
-      "name": "MLX CUDA 12",
-      "targets": [ "mlx", "mlxc" ],
-      "configurePreset": "MLX CUDA 12"
-    },
-    {
-      "name": "MLX CUDA 13",
-      "targets": [ "mlx", "mlxc" ],
-      "configurePreset": "MLX CUDA 13"
+      "name": "MLX Metal",
+      "targets": [ "ollama-mlx-backends" ],
+      "configurePreset": "MLX Metal"
    }
  ]
 }
@@ -2,7 +2,7 @@

 ARG FLAVOR=${TARGETARCH}

-ARG ROCMVERSION=7.2
+ARG ROCMVERSION=7.2.1
 ARG JETPACK5VERSION=r35.4.1
 ARG JETPACK6VERSION=r36.4.0
 ARG CMAKEVERSION=3.31.2
@@ -37,113 +37,171 @@ RUN dnf install -y unzip \
 ENV CMAKE_GENERATOR=Ninja
 ENV LDFLAGS=-s

-FROM base AS cpu
+#
+# GPU toolchain stages — provide compilers for llama-server GPU builds
+#
+
+FROM base AS cpu-deps
 RUN dnf install -y gcc-toolset-11-gcc gcc-toolset-11-gcc-c++
 ENV PATH=/opt/rh/gcc-toolset-11/root/usr/bin:$PATH
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CPU' \
-        && cmake --build --preset 'CPU' -- -l $(nproc) \
-        && cmake --install build --component CPU --strip

-FROM base AS cuda-11
-ARG CUDA11VERSION=11.8
-RUN dnf install -y cuda-toolkit-${CUDA11VERSION//./-}
-ENV PATH=/usr/local/cuda-11/bin:$PATH
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 11' \
-        && cmake --build --preset 'CUDA 11' -- -l $(nproc) \
-        && cmake --install build --component CUDA --strip
-
-FROM base AS cuda-12
+FROM base AS cuda-12-deps
 ARG CUDA12VERSION=12.8
 RUN dnf install -y cuda-toolkit-${CUDA12VERSION//./-}
 ENV PATH=/usr/local/cuda-12/bin:$PATH
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 12' \
-        && cmake --build --preset 'CUDA 12' -- -l $(nproc) \
-        && cmake --install build --component CUDA --strip

-
-FROM base AS cuda-13
+FROM base AS cuda-13-deps
 ARG CUDA13VERSION=13.0
 RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-}
 ENV PATH=/usr/local/cuda-13/bin:$PATH
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 13' \
-        && cmake --build --preset 'CUDA 13' -- -l $(nproc) \
-        && cmake --install build --component CUDA --strip

+FROM base AS rocm-7-deps
+ENV PATH=/opt/rocm/llvm/bin:/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:$PATH

-FROM base AS rocm-7
-ENV PATH=/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:/opt/rocm/hcc/bin:$PATH
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'ROCm 7' \
-        && cmake --build --preset 'ROCm 7' -- -l $(nproc) \
-        && cmake --install build --component HIP --strip
-RUN rm -f dist/lib/ollama/rocm/rocblas/library/*gfx90[06]*
-
-FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK5VERSION} AS jetpack-5
-ARG CMAKEVERSION
-ARG NINJAVERSION
-RUN apt-get update && apt-get install -y curl ccache unzip \
-    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1 \
-    && curl -fsSL -o /tmp/ninja.zip https://github.com/ninja-build/ninja/releases/download/v${NINJAVERSION}/ninja-linux-aarch64.zip \
-    && unzip /tmp/ninja.zip -d /usr/local/bin \
-    && rm /tmp/ninja.zip
-ENV CMAKE_GENERATOR=Ninja
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'JetPack 5' \
-        && cmake --build --preset 'JetPack 5' -- -l $(nproc) \
-        && cmake --install build --component CUDA --strip
-
-FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK6VERSION} AS jetpack-6
-ARG CMAKEVERSION
-ARG NINJAVERSION
-RUN apt-get update && apt-get install -y curl ccache unzip \
-    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1 \
-    && curl -fsSL -o /tmp/ninja.zip https://github.com/ninja-build/ninja/releases/download/v${NINJAVERSION}/ninja-linux-aarch64.zip \
-    && unzip /tmp/ninja.zip -d /usr/local/bin \
-    && rm /tmp/ninja.zip
-ENV CMAKE_GENERATOR=Ninja
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'JetPack 6' \
-        && cmake --build --preset 'JetPack 6' -- -l $(nproc) \
-        && cmake --install build --component CUDA --strip
-
-FROM base AS vulkan
+FROM base AS vulkan-deps
 ARG VULKANVERSION
 RUN ln -s /usr/bin/python3 /usr/bin/python \
    && wget https://sdk.lunarg.com/sdk/download/${VULKANVERSION}/linux/vulkansdk-linux-x86_64-${VULKANVERSION}.tar.xz -O /tmp/vulkansdk.tar.xz \
    && tar xvf /tmp/vulkansdk.tar.xz -C /tmp \
    && /tmp/${VULKANVERSION}/vulkansdk -j 8 vulkan-headers \
+    && /tmp/${VULKANVERSION}/vulkansdk -j 8 spirv-headers \
    && /tmp/${VULKANVERSION}/vulkansdk -j 8 shaderc \
    && cp -r /tmp/${VULKANVERSION}/x86_64/include/* /usr/local/include/ \
    && cp -r /tmp/${VULKANVERSION}/x86_64/lib/* /usr/local/lib \
+    && cp -r /tmp/${VULKANVERSION}/x86_64/share/* /usr/local/share/ \
    && cp -r /tmp/${VULKANVERSION}/x86_64/bin/* /usr/local/bin/ \
    && rm -rf /tmp/${VULKANVERSION} /tmp/vulkansdk.tar.xz
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+ENV VULKAN_SDK=/usr/local
+
+#
+# llama-server stages — rebuild when LLAMA_CPP_VERSION, llama/server/, or llama/compat/ changes.
+#
+# CPU stage: llama-server + ggml-base + ggml-cpu variants → lib/ollama/
+# GPU stages: GPU backend .so only → lib/ollama/<variant>/
+#
+
+FROM cpu-deps AS llama-server-cpu
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
 RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'Vulkan' \
-        && cmake --build --preset 'Vulkan' -- -l $(nproc) \
-        && cmake --install build --component Vulkan --strip
+    cmake -S llama/server --preset cpu \
+        && cmake --build build/llama-server-cpu -- -l $(nproc) \
+        && cmake --install build/llama-server-cpu --component llama-server --strip \
+        && for lib in \
+            /usr/lib64/libgomp.so* \
+            /usr/lib64/libomp.so* \
+            /opt/rh/gcc-toolset-11/root/usr/lib64/libgomp.so* \
+            /opt/rh/gcc-toolset-11/root/usr/lib64/libomp.so*; do \
+                [ -e "$lib" ] && cp -a "$lib" dist/lib/ollama/ || true; \
+            done
+
+FROM scratch AS publish-llama-server-cpu
+COPY --from=llama-server-cpu dist/lib/ollama /lib/ollama/
+
+FROM cuda-12-deps AS llama-server-cuda_v12
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset llama_cuda_v12_linux \
+        && cmake --build build/llama-server-cuda_v12 -- -l $(nproc) \
+        && cmake --install build/llama-server-cuda_v12 --component llama-server --strip
+
+FROM scratch AS publish-llama-server-cuda_v12
+COPY --from=llama-server-cuda_v12 dist/lib/ollama /lib/ollama/
+
+FROM cuda-13-deps AS llama-server-cuda_v13
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset llama_cuda_v13_linux \
+        && cmake --build build/llama-server-cuda_v13 -- -l $(nproc) \
+        && cmake --install build/llama-server-cuda_v13 --component llama-server --strip
+
+FROM scratch AS publish-llama-server-cuda_v13
+COPY --from=llama-server-cuda_v13 dist/lib/ollama /lib/ollama/
+
+FROM rocm-7-deps AS llama-server-rocm_v7_2
+ENV CC=clang CXX=clang++
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset rocm_v7_2_linux \
+        && cmake --build build/llama-server-rocm_v7_2 -- -l $(nproc) \
+        && cmake --install build/llama-server-rocm_v7_2 --component llama-server --strip
+RUN rm -f dist/lib/ollama/rocm_v7_2/rocblas/library/*gfx90[06]*
+
+FROM scratch AS publish-llama-server-rocm_v7_2
+COPY --from=llama-server-rocm_v7_2 dist/lib/ollama /lib/ollama/
+
+FROM vulkan-deps AS llama-server-vulkan
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset vulkan \
+        && cmake --build build/llama-server-vulkan -- -l $(nproc) \
+        && cmake --install build/llama-server-vulkan --component llama-server --strip
+
+FROM scratch AS publish-llama-server-vulkan
+COPY --from=llama-server-vulkan dist/lib/ollama /lib/ollama/
+
+#
+# JetPack stages — self-contained with their own base images
+#
+
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK5VERSION} AS jetpack-5
+ARG CMAKEVERSION
+ARG NINJAVERSION
+RUN apt-get update && apt-get install -y curl ccache git unzip \
+    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1 \
+    && curl -fsSL -o /tmp/ninja.zip https://github.com/ninja-build/ninja/releases/download/v${NINJAVERSION}/ninja-linux-aarch64.zip \
+    && unzip /tmp/ninja.zip -d /usr/local/bin \
+    && rm /tmp/ninja.zip
+ENV CMAKE_GENERATOR=Ninja
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset llama_cuda_jetpack5 \
+        && cmake --build build/llama-server-cuda_jetpack5 -- -l $(nproc) \
+        && cmake --install build/llama-server-cuda_jetpack5 --component llama-server --strip
+
+FROM scratch AS publish-llama-server-cuda_jetpack5
+COPY --from=jetpack-5 dist/lib/ollama /lib/ollama/
+
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK6VERSION} AS jetpack-6
+ARG CMAKEVERSION
+ARG NINJAVERSION
+RUN apt-get update && apt-get install -y curl ccache git unzip \
+    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1 \
+    && curl -fsSL -o /tmp/ninja.zip https://github.com/ninja-build/ninja/releases/download/v${NINJAVERSION}/ninja-linux-aarch64.zip \
+    && unzip /tmp/ninja.zip -d /usr/local/bin \
+    && rm /tmp/ninja.zip
+ENV CMAKE_GENERATOR=Ninja
+COPY LLAMA_CPP_VERSION .
+COPY llama/server llama/server
+COPY llama/compat llama/compat
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake -S llama/server --preset llama_cuda_jetpack6 \
+        && cmake --build build/llama-server-cuda_jetpack6 -- -l $(nproc) \
+        && cmake --install build/llama-server-cuda_jetpack6 --component llama-server --strip
+
+FROM scratch AS publish-llama-server-cuda_jetpack6
+COPY --from=jetpack-6 dist/lib/ollama /lib/ollama/
+
+#
+# MLX stage
+#

 FROM base AS mlx
 ARG CUDA13VERSION=13.0
+ARG OLLAMA_MLX_BUILD_JOBS=
+ARG OLLAMA_MLX_NVCC_THREADS=2
+ARG MLX_CUDA_RAM_MB=
 RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-} \
    && dnf install -y openblas-devel lapack-devel \
    && dnf install -y libcudnn9-cuda-13 libcudnn9-devel-cuda-13 \
@@ -154,7 +212,7 @@ ENV LAPACK_INCLUDE_DIRS=/usr/include/openblas
 ENV CGO_LDFLAGS="-L/usr/local/cuda-13/lib64 -L/usr/local/cuda-13/targets/x86_64-linux/lib/stubs"
 WORKDIR /go/src/github.com/ollama/ollama
 COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+COPY cmake cmake
 COPY x/imagegen/mlx x/imagegen/mlx
 COPY go.mod go.sum .
 COPY MLX_VERSION MLX_C_VERSION .
@@ -170,9 +228,15 @@ RUN --mount=type=cache,target=/root/.ccache \
    && if [ -f /tmp/local-mlx-c/CMakeLists.txt ]; then \
        export OLLAMA_MLX_C_SOURCE=/tmp/local-mlx-c; \
    fi \
-    && cmake --preset 'MLX CUDA 13' -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas \
-        && cmake --build --preset 'MLX CUDA 13' -- -l $(nproc) \
-        && cmake --install build --component MLX --strip
+    && cmake -S . -B build/mlx_cuda_v13 -DOLLAMA_MLX_BACKENDS=cuda_v13 -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas -DCMAKE_CUDA_FLAGS="-t ${OLLAMA_MLX_NVCC_THREADS}" ${MLX_CUDA_RAM_MB:+-DMLX_CUDA_RAM_MB=${MLX_CUDA_RAM_MB}} -DOLLAMA_PAYLOAD_INSTALL_PREFIX=/go/src/github.com/ollama/ollama/dist \
+        && cmake --build build/mlx_cuda_v13 --target ollama-mlx-cuda_v13 -- -l $(nproc) ${OLLAMA_MLX_BUILD_JOBS:+-j ${OLLAMA_MLX_BUILD_JOBS}}
+
+FROM scratch AS publish-mlx
+COPY --from=mlx /go/src/github.com/ollama/ollama/dist/lib/ollama /lib/ollama/
+
+#
+# Go build
+#

 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -190,35 +254,62 @@ ENV CGO_CXXFLAGS="${CGO_CXXFLAGS}"
 RUN --mount=type=cache,target=/root/.cache/go-build \
    go build -trimpath -buildmode=pie -o /bin/ollama .

+FROM scratch AS publish-go
+COPY --from=build /bin/ollama /bin/ollama
+
+#
+# Assembly stages — combine llama-server variants + GPU runtime libs
+#
+
 FROM --platform=linux/amd64 scratch AS amd64
-# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
-COPY --from=cuda-13 dist/lib/ollama /lib/ollama/
-COPY --from=vulkan  dist/lib/ollama  /lib/ollama/
+COPY --from=llama-server-cpu      dist/lib/ollama /lib/ollama/
+COPY --from=llama-server-cuda_v12 dist/lib/ollama /lib/ollama/
+COPY --from=llama-server-cuda_v13 dist/lib/ollama /lib/ollama/
+COPY --from=llama-server-vulkan   dist/lib/ollama /lib/ollama/
 COPY --from=mlx     /go/src/github.com/ollama/ollama/dist/lib/ollama /lib/ollama/

 FROM --platform=linux/arm64 scratch AS arm64
-# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
-COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/
+COPY --from=llama-server-cpu dist/lib/ollama /lib/ollama/
+COPY --from=llama-server-cuda_v12 dist/lib/ollama /lib/ollama/
+COPY --from=llama-server-cuda_v13 dist/lib/ollama /lib/ollama/
 COPY --from=jetpack-5 dist/lib/ollama/ /lib/ollama/
 COPY --from=jetpack-6 dist/lib/ollama/ /lib/ollama/

 FROM scratch AS rocm
-COPY --from=rocm-7 dist/lib/ollama /lib/ollama
+COPY --from=llama-server-cpu  dist/lib/ollama /lib/ollama
+COPY --from=llama-server-rocm_v7_2 dist/lib/ollama /lib/ollama

-FROM ${FLAVOR} AS archive
-COPY --from=cpu dist/lib/ollama /lib/ollama
+FROM --platform=linux/amd64 scratch AS amd64-archive
+COPY --from=amd64 /lib/ollama /lib/ollama/
+COPY --from=llama-server-rocm_v7_2 dist/lib/ollama /lib/ollama/
+
+FROM --platform=linux/arm64 scratch AS arm64-archive
+COPY --from=arm64 /lib/ollama /lib/ollama/
+
+FROM ${TARGETARCH}-archive AS archive
+COPY --from=build /bin/ollama /bin/ollama
+
+FROM ${FLAVOR} AS image-archive
 COPY --from=build /bin/ollama /bin/ollama

 FROM ubuntu:24.04
-RUN apt-get update \
+ARG APT_MIRROR=http://archive.ubuntu.com/ubuntu
+ARG APT_PORTS_MIRROR=http://ports.ubuntu.com/ubuntu-ports
+RUN sed -i \
+        -e "s|http://archive.ubuntu.com/ubuntu|$APT_MIRROR|g" \
+        -e "s|http://ports.ubuntu.com/ubuntu-ports|$APT_PORTS_MIRROR|g" \
+        /etc/apt/sources.list.d/ubuntu.sources \
+    && apt-get update \
    && apt-get install -y ca-certificates libvulkan1 libopenblas0 \
+    && sed -i \
+        -e "s|$APT_MIRROR|http://archive.ubuntu.com/ubuntu|g" \
+        -e "s|$APT_PORTS_MIRROR|http://ports.ubuntu.com/ubuntu-ports|g" \
+        /etc/apt/sources.list.d/ubuntu.sources \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/*
-COPY --from=archive /bin /usr/bin
+COPY --from=image-archive /bin /usr/bin
 ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
-COPY --from=archive /lib/ollama /usr/lib/ollama
+COPY --from=image-archive /lib/ollama /usr/lib/ollama
 ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
 ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
 ENV NVIDIA_VISIBLE_DEVICES=all
@@ -0,0 +1 @@
+b9888
@@ -1 +1 @@
-0726ca922fc902c4c61ef9c27d94132be418e945
+fba4470b89073180056c9ea46c443051375f7399
@@ -1 +1 @@
-38ad257088fb2193ad47e527cf6534a689f30943
+de7b4ed986b6d6f55b8ace5e73c24d1ca0bea89b
@@ -1,76 +0,0 @@
-UPSTREAM=https://github.com/ggml-org/llama.cpp.git
-WORKDIR=llama/vendor
-FETCH_HEAD=ec98e2002
-
-.PHONY: help
-help:
-	@echo "Available targets:"
-	@echo "    sync                 Sync with upstream repositories"
-	@echo "    checkout             Checkout upstream repository"
-	@echo "    apply-patches        Apply patches to local repository"
-	@echo "    format-patches       Format patches from local repository"
-	@echo "    clean                Clean local repository"
-	@echo
-	@echo "Example:"
-	@echo "    make -f $(lastword $(MAKEFILE_LIST)) clean apply-patches sync"
-
-.PHONY: sync
-sync: llama/build-info.cpp ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal
-
-llama/build-info.cpp: llama/build-info.cpp.in llama/llama.cpp
-	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' <$< >$@
-
-ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal: ml/backend/ggml/ggml
-	go generate ./$(@D)
-
-.PHONY: llama/llama.cpp
-llama/llama.cpp: llama/vendor
-	rsync -arvzc --delete -f "include LICENSE" -f "merge $@/.rsync-filter" $(addprefix $<,/LICENSE /) $@
-
-.PHONY: ml/backend/ggml/ggml
-ml/backend/ggml/ggml: llama/vendor
-	rsync -arvzc --delete -f "include LICENSE" -f "merge $@/.rsync-filter" $(addprefix $<,/LICENSE /ggml/) $@
-
-PATCHES=$(wildcard llama/patches/*.patch)
-PATCHED=$(join $(dir $(PATCHES)), $(addsuffix ed, $(addprefix ., $(notdir $(PATCHES)))))
-
-.PHONY: apply-patches
-.NOTPARALLEL:
-apply-patches: $(PATCHED)
-
-llama/patches/.%.patched: llama/patches/%.patch
-	@if git -c user.name=nobody -c 'user.email=<>' -C $(WORKDIR) am -3 $(realpath $<); then \
-		touch $@;                                                                           \
-	else                                                                                    \
-		echo "Patch failed. Resolve any conflicts then continue.";                          \
-		echo "1. Run 'git -C $(WORKDIR) am --continue'";                                    \
-		echo "2. Run 'make -f $(lastword $(MAKEFILE_LIST)) format-patches'";                \
-		echo "3. Run 'make -f $(lastword $(MAKEFILE_LIST)) clean apply-patches'";           \
-		exit 1;                                                                             \
-	fi
-
-.PHONY: checkout
-checkout: $(WORKDIR)
-	git -C $(WORKDIR) fetch
-	git -C $(WORKDIR) checkout -f $(FETCH_HEAD)
-
-$(WORKDIR):
-	git clone $(UPSTREAM) $(WORKDIR)
-
-.PHONY: format-patches
-format-patches: llama/patches
-	git -C $(WORKDIR) format-patch \
-		--no-signature \
-		--no-numbered \
-		--zero-commit \
-		-o $(realpath $<) \
-		$(FETCH_HEAD)
-
-.PHONY: clean
-clean: checkout
-	@git -C $(WORKDIR) am --abort || true
-	$(RM) llama/patches/.*.patched
-
-.PHONY: print-base
-print-base:
-	@echo $(FETCH_HEAD)
@@ -55,7 +55,7 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 ollama
 ```

-You'll be prompted to run a model or connect Ollama to your existing agents or applications such as `claude`, `codex`, `openclaw` and more.
+You'll be prompted to run a model or connect Ollama to your existing agents or applications such as `Claude Code`, `OpenClaw`, `OpenCode` , `Codex`, `Copilot`,  and more.

 ### Coding

@@ -65,7 +65,7 @@ To launch a specific integration:
 ollama launch claude
 ```

-Supported integrations include [Claude Code](https://docs.ollama.com/integrations/claude-code), [Codex](https://docs.ollama.com/integrations/codex), [Droid](https://docs.ollama.com/integrations/droid), and [OpenCode](https://docs.ollama.com/integrations/opencode).
+Supported integrations include [Claude Code](https://docs.ollama.com/integrations/claude-code), [Codex](https://docs.ollama.com/integrations/codex), [Copilot CLI](https://docs.ollama.com/integrations/copilot-cli), [Droid](https://docs.ollama.com/integrations/droid), and [OpenCode](https://docs.ollama.com/integrations/opencode).

 ### AI assistant

@@ -77,10 +77,10 @@ ollama launch openclaw

 ### Chat with a model

-Run and chat with [Gemma 3](https://ollama.com/library/gemma3):
+Run and chat with [Gemma 4](https://ollama.com/library/gemma4):

 ```
-ollama run gemma3
+ollama run gemma4
 ```

 See [ollama.com/library](https://ollama.com/library) for the full list.
@@ -93,7 +93,7 @@ Ollama has a REST API for running and managing models.

 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "gemma3",
+  "model": "gemma4",
  "messages": [{
    "role": "user",
    "content": "Why is the sky blue?"
@@ -113,7 +113,7 @@ pip install ollama
 ```python
 from ollama import chat

-response = chat(model='gemma3', messages=[
+response = chat(model='gemma4', messages=[
  {
    'role': 'user',
    'content': 'Why is the sky blue?',
@@ -132,7 +132,7 @@ npm i ollama
 import ollama from "ollama";

 const response = await ollama.chat({
-  model: "gemma3",
+  model: "gemma4",
  messages: [{ role: "user", content: "Why is the sky blue?" }],
 });
 console.log(response.message.content);
@@ -0,0 +1,209 @@
+package agent
+
+import (
+	"context"
+	"strings"
+	"sync"
+)
+
+type ApprovalRequest struct {
+	WorkingDir string
+	Calls      []ApprovalToolCall
+}
+
+func (r *ApprovalRequest) AddToolCall(id, name string, args map[string]any) {
+	r.Calls = append(r.Calls, ApprovalToolCall{
+		ToolCallID:    id,
+		ToolName:      name,
+		Args:          args,
+		ApprovalScope: toolApprovalScope(name, args),
+	})
+}
+
+type ApprovalToolCall struct {
+	ToolCallID    string
+	ToolName      string
+	Args          map[string]any
+	ApprovalScope string
+}
+
+type Approval struct {
+	Allow       bool
+	AllowAll    bool
+	AllowScopes []string
+	Reason      string
+}
+
+type ApprovalPrompter interface {
+	PromptApproval(context.Context, ApprovalRequest) (Approval, error)
+}
+
+type ApprovalState struct {
+	mu       sync.RWMutex
+	allowAll bool
+	scopes   map[string]bool
+}
+
+func (s *ApprovalState) Set(allowAll bool, scopes map[string]bool) {
+	if s == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.allowAll = allowAll
+	s.scopes = cloneApprovalScopes(scopes)
+}
+
+func (s *ApprovalState) SetAllowAll(allowAll bool) {
+	if s == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.allowAll = allowAll
+}
+
+func (s *ApprovalState) AllowAll() bool {
+	if s == nil {
+		return false
+	}
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	return s.allowAll
+}
+
+func (s *ApprovalState) Allows(scope string) bool {
+	if s == nil {
+		return false
+	}
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	return s.allowAll || s.scopes[scope]
+}
+
+func (s *ApprovalState) Apply(result *Approval) {
+	if s == nil || result == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if result.AllowAll {
+		result.Allow = true
+		s.allowAll = true
+	}
+	if len(result.AllowScopes) > 0 {
+		result.Allow = true
+		if s.scopes == nil {
+			s.scopes = make(map[string]bool, len(result.AllowScopes))
+		}
+		for _, scope := range result.AllowScopes {
+			scope = strings.TrimSpace(scope)
+			if scope != "" {
+				s.scopes[scope] = true
+			}
+		}
+	}
+}
+
+func (s *ApprovalState) AllowScopes(scopes []string) {
+	if s == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.scopes == nil {
+		s.scopes = make(map[string]bool, len(scopes))
+	}
+	for _, scope := range scopes {
+		scope = strings.TrimSpace(scope)
+		if scope != "" {
+			s.scopes[scope] = true
+		}
+	}
+}
+
+func cloneApprovalScopes(src map[string]bool) map[string]bool {
+	if len(src) == 0 {
+		return nil
+	}
+	dst := make(map[string]bool, len(src))
+	for scope, allowed := range src {
+		if allowed {
+			dst[scope] = true
+		}
+	}
+	return dst
+}
+
+func (s *Session) needsApproval(tool Tool, name string, args map[string]any) bool {
+	return ToolRequiresApproval(tool, args) && !s.allows(toolApprovalScope(name, args))
+}
+
+// allows reports whether scope is permitted by the session's accumulated approval state.
+func (s *Session) allows(scope string) bool {
+	if s == nil || s.ApprovalState == nil {
+		return false
+	}
+	return s.ApprovalState.Allows(scope)
+}
+
+// applyApproval merges an approval result into the session's state and marks
+// the result as allowed when scopes or allow-all were granted. It mutates
+// result.Allow so the caller can branch on the effective decision.
+func (s *Session) applyApproval(result *Approval) {
+	if s == nil || result == nil {
+		return
+	}
+	if s.ApprovalState == nil {
+		s.ApprovalState = &ApprovalState{}
+	}
+	s.ApprovalState.Apply(result)
+}
+
+func (s *Session) authorizeToolCalls(ctx context.Context, req ApprovalRequest) (Approval, error) {
+	if s == nil || len(req.Calls) == 0 || (s.ApprovalState != nil && s.ApprovalState.AllowAll()) {
+		return Approval{Allow: true}, nil
+	}
+	if s.ApprovalPrompter == nil {
+		return Approval{
+			Reason: "Tool execution requires approval, but no approval prompter is available.",
+		}, nil
+	}
+
+	result, err := s.ApprovalPrompter.PromptApproval(ctx, req)
+	if err != nil {
+		return Approval{}, err
+	}
+	s.applyApproval(&result)
+	return result, nil
+}
+
+// toolApprovalScope returns the approval scope key for a tool invocation.
+//
+// For shell tools (bash/powershell) the scope is "<tool>\x00<command>": the
+// exact, trimmed command byte string. "Always allow this command" therefore
+// matches ONLY that precise string — any whitespace, quoting, or casing
+// variant, or any command that is a superset of the approved one, will
+// re-prompt. The NUL separator is safe because a shell command string cannot
+// contain a literal NUL. For all other tools the scope is the tool name.
+func toolApprovalScope(toolName string, args map[string]any) string {
+	toolName = strings.TrimSpace(toolName)
+	if isShellApprovalTool(toolName) {
+		if command, ok := stringArg(args, "command"); ok {
+			command = strings.TrimSpace(command)
+			if command != "" {
+				return toolName + "\x00" + command
+			}
+		}
+	}
+	return toolName
+}
+
+func isShellApprovalTool(name string) bool {
+	return name == "bash" || name == "powershell"
+}
+
+func stringArg(args map[string]any, key string) (string, bool) {
+	value, ok := args[key].(string)
+	return value, ok
+}
@@ -0,0 +1,55 @@
+package agent
+
+import "testing"
+
+func TestApprovalRequestScopesShellCommandsToExactCommand(t *testing.T) {
+	req := ApprovalRequest{}
+	req.AddToolCall("call-1", "bash", map[string]any{"command": " pwd "})
+	req.AddToolCall("call-2", "powershell", map[string]any{"command": "Get-ChildItem"})
+	req.AddToolCall("call-3", "edit", map[string]any{"path": "README.md"})
+
+	tests := []struct {
+		index int
+		want  string
+	}{
+		{index: 0, want: "bash\x00pwd"},
+		{index: 1, want: "powershell\x00Get-ChildItem"},
+		{index: 2, want: "edit"},
+	}
+	for _, tt := range tests {
+		if got := req.Calls[tt.index].ApprovalScope; got != tt.want {
+			t.Fatalf("call %d approval scope = %q, want %q", tt.index, got, tt.want)
+		}
+	}
+}
+
+func TestSessionApplyApprovalScopes(t *testing.T) {
+	session := &Session{}
+	result := Approval{AllowScopes: []string{"edit", "bash\x00pwd", " "}}
+
+	session.applyApproval(&result)
+
+	if !result.Allow {
+		t.Fatal("scoped approval should allow the current request")
+	}
+	if !session.allows("edit") || !session.allows("bash\x00pwd") {
+		t.Fatal("scoped approval was not saved")
+	}
+	if session.allows("bash") || session.allows("bash\x00ls") {
+		t.Fatal("shell approval was too broad")
+	}
+	if session.ApprovalState.AllowAll() {
+		t.Fatal("allow all = true, want false for scoped approval")
+	}
+}
+
+func TestSessionApplyApprovalAllowAll(t *testing.T) {
+	session := &Session{}
+	result := Approval{AllowAll: true}
+
+	session.applyApproval(&result)
+
+	if !result.Allow || !session.allows("anything") {
+		t.Fatalf("allow all = %v result = %#v, want allow all", session.ApprovalState.AllowAll(), result)
+	}
+}
@@ -0,0 +1,629 @@
+package agent
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Compaction wire-format. These constants and helpers are the single canonical
+// definition of how a compacted turn is represented in message history.
+const (
+	CompactionSummaryMessagePrefix = "Conversation summary:\n"
+	CompactionToolName             = "summary"
+	CompactionToolCallID           = "ollama_compaction"
+	CompactionContinueInstruction  = "continue the task in progress. the history has been compacted, do not mention compaction to the user"
+)
+
+const (
+	defaultCompactionContextWindowTokens = 32768
+	defaultCompactionKeepUserTurns       = 3
+	defaultCompactionThreshold           = 0.8
+	compactOnlySummaryContextTokens      = 16000
+
+	maxCompactionSummaryBytes  = 16 * 1024
+	compactionSummaryTruncated = "\n\n[summary truncated]"
+
+	compactionSystemPrompt = "Summarize the archived part of an Ollama agent conversation. Preserve user goals, decisions, files, commands, tool results, and unresolved tasks needed to continue. Omit private reasoning and return only the summary."
+)
+
+type Compactor interface {
+	MaybeCompact(context.Context, CompactionRequest) (CompactionResult, error)
+}
+
+type CompactionOptions struct {
+	ContextWindowTokens int
+	KeepUserTurns       int
+	Threshold           float64
+}
+
+type CompactionRequest struct {
+	ChatID        string
+	Model         string
+	SystemPrompt  string
+	Messages      []api.Message
+	Tools         api.Tools
+	Format        string
+	Latest        api.ChatResponse
+	Options       map[string]any
+	KeepAlive     *api.Duration
+	Think         *api.ThinkValue
+	Force         bool
+	ContinueTask  bool
+	KeepUserTurns *int
+	Progress      func(CompactionProgress)
+}
+
+type CompactionProgress struct {
+	Tokens int
+}
+
+type CompactionResult struct {
+	Messages  []api.Message
+	Compacted bool
+	Due       bool
+	Summary   string
+	Reason    string
+}
+
+type SimpleCompactor struct {
+	Client  ChatClient
+	Options CompactionOptions
+}
+
+func (c *SimpleCompactor) MaybeCompact(ctx context.Context, req CompactionRequest) (CompactionResult, error) {
+	result := CompactionResult{Messages: req.Messages}
+	if c == nil {
+		return result, nil
+	}
+
+	result.Due = req.Force || c.shouldCompact(req)
+	if !result.Due {
+		return result, nil
+	}
+	if c.Client == nil {
+		result.Reason = "compaction is unavailable"
+		return result, nil
+	}
+
+	keepUserTurns := c.keepUserTurns(req.Options)
+	if req.KeepUserTurns != nil {
+		keepUserTurns = *req.KeepUserTurns
+	}
+	prefix, previousSummary, archive, suffix, _, ok := splitCompactionMessages(req.Messages, keepUserTurns)
+	if !ok || len(archive) == 0 {
+		result.Reason = "nothing to compact"
+		return result, nil
+	}
+
+	summary, err := c.summarize(ctx, req, previousSummary, archive)
+	if err != nil {
+		result.Reason = err.Error()
+		return result, err
+	}
+	summary = truncateCompactionSummary(strings.TrimSpace(summary))
+	if summary == "" {
+		summary, err = c.summarizeEmptyFallback(ctx, req, previousSummary, archive)
+		if err != nil {
+			result.Reason = err.Error()
+			return result, err
+		}
+		summary = truncateCompactionSummary(strings.TrimSpace(summary))
+	}
+	if summary == "" {
+		result.Reason = "summary was empty"
+		return result, nil
+	}
+
+	compacted := make([]api.Message, 0, len(prefix)+len(suffix)+2)
+	compacted = append(compacted, prefix...)
+	compacted = append(compacted, CompactionSummaryMessages(summary, req.ContinueTask)...)
+	compacted = append(compacted, suffix...)
+	result.Messages = compacted
+	result.Compacted = true
+	result.Summary = summary
+	return result, nil
+}
+
+func (c *SimpleCompactor) shouldCompact(req CompactionRequest) bool {
+	contextWindow := c.contextWindowTokens(req.Options)
+	threshold := int(float64(contextWindow) * c.threshold())
+	if threshold <= 0 {
+		return false
+	}
+	if req.Latest.PromptEvalCount > 0 && req.Latest.PromptEvalCount >= threshold {
+		return true
+	}
+	return estimateCompactionRequestTokens(req) >= threshold
+}
+
+func (c *SimpleCompactor) contextWindowTokens(options map[string]any) int {
+	return ResolveContextWindowTokens(options, c.Options.ContextWindowTokens)
+}
+
+func (c *SimpleCompactor) keepUserTurns(options map[string]any) int {
+	contextWindow := c.contextWindowTokens(options)
+	if contextWindow > 0 && contextWindow < compactOnlySummaryContextTokens {
+		return 0
+	}
+	if c.Options.KeepUserTurns > 0 {
+		return c.Options.KeepUserTurns
+	}
+	return defaultCompactionKeepUserTurns
+}
+
+func (c *SimpleCompactor) threshold() float64 {
+	return ResolveCompactionThreshold(c.Options.Threshold)
+}
+
+func ResolveContextWindowTokens(options map[string]any, configured int) int {
+	if n := intOption(options, "num_ctx"); n > 0 {
+		return n
+	}
+	if configured > 0 {
+		return configured
+	}
+	return defaultCompactionContextWindowTokens
+}
+
+func ResolveCompactionThreshold(configured float64) float64 {
+	if configured > 0 {
+		return configured
+	}
+	return defaultCompactionThreshold
+}
+
+func (c *SimpleCompactor) summarize(ctx context.Context, req CompactionRequest, previousSummary string, archive []api.Message) (string, error) {
+	body, err := compactionPrompt(previousSummary, archive, c.compactionPromptBodyBudgetTokens(req.Options))
+	if err != nil {
+		return "", err
+	}
+
+	chatReq := &api.ChatRequest{
+		Model: req.Model,
+		Messages: []api.Message{
+			{
+				Role:    "system",
+				Content: compactionSystemPrompt,
+			},
+			{
+				Role:    "user",
+				Content: body,
+			},
+		},
+		Options: req.Options,
+		Think:   req.Think,
+	}
+	if req.KeepAlive != nil {
+		chatReq.KeepAlive = req.KeepAlive
+	}
+
+	var summary strings.Builder
+	if err := c.Client.Chat(ctx, chatReq, func(response api.ChatResponse) error {
+		summary.WriteString(response.Message.Content)
+		if req.Progress != nil {
+			tokens := response.EvalCount
+			if tokens <= 0 {
+				tokens = estimateCompactionTokens(summary.String())
+			}
+			req.Progress(CompactionProgress{Tokens: tokens})
+		}
+		return nil
+	}); err != nil {
+		return "", err
+	}
+	return summary.String(), nil
+}
+
+func (c *SimpleCompactor) summarizeEmptyFallback(ctx context.Context, req CompactionRequest, previousSummary string, archive []api.Message) (string, error) {
+	retry := req
+	retry.Think = &api.ThinkValue{Value: false}
+	summary, err := c.summarize(ctx, retry, previousSummary, archive)
+	if err == nil {
+		return summary, nil
+	}
+	if !isUnsupportedCompactionThinkError(err) {
+		return "", err
+	}
+	if req.Think == nil {
+		return "", nil
+	}
+	retry.Think = nil
+	return c.summarize(ctx, retry, previousSummary, archive)
+}
+
+func isUnsupportedCompactionThinkError(err error) bool {
+	if err == nil {
+		return false
+	}
+	text := strings.ToLower(err.Error())
+	if !strings.Contains(text, "think") {
+		return false
+	}
+	var statusErr api.StatusError
+	if errors.As(err, &statusErr) && statusErr.StatusCode != 0 {
+		return statusErr.StatusCode == http.StatusBadRequest
+	}
+	return strings.Contains(text, "does not support") || strings.Contains(text, "not supported") || strings.Contains(text, "unsupported")
+}
+
+// compactionSummaryMessageForTask renders a compaction summary as the content
+// string stored on the synthetic tool-result message.
+func compactionSummaryMessageForTask(summary string, continueTask bool) string {
+	content := CompactionSummaryMessagePrefix + strings.TrimSpace(summary)
+	if continueTask {
+		content = strings.TrimSpace(content) + "\n\n" + CompactionContinueInstruction
+	}
+	return content
+}
+
+// CompactionSummaryMessages renders a compaction summary as the assistant
+// tool-call plus tool-result pair that represents a compacted turn in the
+// message history.
+func CompactionSummaryMessages(summary string, continueTask bool) []api.Message {
+	return []api.Message{
+		{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{{
+				ID: CompactionToolCallID,
+				Function: api.ToolCallFunction{
+					Name: CompactionToolName,
+				},
+			}},
+		},
+		{
+			Role:       "tool",
+			ToolName:   CompactionToolName,
+			ToolCallID: CompactionToolCallID,
+			Content:    compactionSummaryMessageForTask(summary, continueTask),
+		},
+	}
+}
+
+func (c *SimpleCompactor) compactionPromptBodyBudgetTokens(options map[string]any) int {
+	contextWindow := c.contextWindowTokens(options)
+	threshold := int(float64(contextWindow) * c.threshold())
+	if threshold <= 0 {
+		return 0
+	}
+	systemTokens := estimateCompactionTokens("system") + estimateCompactionTokens(compactionSystemPrompt)
+	userRoleTokens := estimateCompactionTokens("user")
+	budget := threshold - systemTokens - userRoleTokens
+	if budget <= 0 {
+		return 0
+	}
+	return budget
+}
+
+func truncateCompactionSummary(summary string) string {
+	if len(summary) <= maxCompactionSummaryBytes {
+		return summary
+	}
+	limit := maxCompactionSummaryBytes - len(compactionSummaryTruncated)
+	if limit < 0 {
+		limit = 0
+	}
+	var b strings.Builder
+	for _, r := range summary {
+		if b.Len()+len(string(r)) > limit {
+			break
+		}
+		b.WriteRune(r)
+	}
+	return strings.TrimSpace(b.String()) + compactionSummaryTruncated
+}
+
+func estimateCompactionTokens(text string) int {
+	text = strings.TrimSpace(text)
+	if text == "" {
+		return 0
+	}
+	return max(1, (len([]rune(text))+3)/4)
+}
+
+func estimateMessagesTokens(messages []api.Message) int {
+	var total int
+	for _, msg := range messages {
+		total += estimateCompactionTokens(msg.Role)
+		total += estimateCompactionTokens(msg.Content)
+		total += estimateCompactionTokens(msg.Thinking)
+		total += estimateCompactionTokens(msg.ToolName)
+		total += estimateCompactionTokens(msg.ToolCallID)
+		for _, call := range msg.ToolCalls {
+			total += estimateCompactionTokens(call.Function.Name)
+			total += estimateCompactionTokens(call.Function.Arguments.String())
+		}
+	}
+	return total
+}
+
+func estimateCompactionRequestTokens(req CompactionRequest) int {
+	requestMessages := sanitizeMessagesForEstimate(req.Messages)
+	if strings.TrimSpace(req.SystemPrompt) != "" {
+		requestMessages = make([]api.Message, 0, len(req.Messages)+1)
+		requestMessages = append(requestMessages, api.Message{Role: "system", Content: strings.TrimSpace(req.SystemPrompt)})
+		requestMessages = append(requestMessages, sanitizeMessagesForEstimate(req.Messages)...)
+	}
+
+	payload := struct {
+		Messages []api.Message   `json:"messages,omitempty"`
+		Tools    api.Tools       `json:"tools,omitempty"`
+		Format   json.RawMessage `json:"format,omitempty"`
+	}{
+		Messages: requestMessages,
+		Tools:    req.Tools,
+	}
+	if rawFormat, ok := compactionFormatForEstimate(req.Format); ok {
+		payload.Format = rawFormat
+	}
+	if data, err := json.Marshal(payload); err == nil {
+		return estimateCompactionTokens(string(data))
+	}
+
+	total := estimateMessagesTokens(requestMessages)
+	total += estimateCompactionTokens(req.Tools.String())
+	total += estimateCompactionTokens(req.Format)
+	return total
+}
+
+func (s *Session) estimateRunPromptTokens(opts RunOptions, messages []api.Message) int {
+	return estimateCompactionRequestTokens(CompactionRequest{
+		SystemPrompt: opts.SystemPrompt,
+		Messages:     messages,
+		Tools:        s.availableTools(),
+		Format:       opts.Format,
+		Options:      opts.Options,
+	})
+}
+
+func (s *Session) checkPreflightPromptBudget(opts RunOptions, messages []api.Message) error {
+	contextWindow := s.contextWindowTokens(opts)
+	if contextWindow <= 0 {
+		return nil
+	}
+	estimated := s.estimateRunPromptTokens(opts, messages)
+	if estimated < contextWindow {
+		return nil
+	}
+	return fmt.Errorf("prompt is too large for the current context (~%d/%d tokens). Reduce the system prompt or message history, compact the conversation, or use a model with a larger context", estimated, contextWindow)
+}
+
+func (s *Session) checkPostCompactionPromptBudget(opts RunOptions, messages []api.Message) error {
+	contextWindow := s.contextWindowTokens(opts)
+	if contextWindow <= 0 {
+		return nil
+	}
+	estimated := s.estimateRunPromptTokens(opts, messages)
+	if estimated < contextWindow {
+		return nil
+	}
+	return fmt.Errorf("history is still too large after compaction (~%d/%d tokens). Start a fresh request, reduce the system prompt or history, or use a model with a larger context", estimated, contextWindow)
+}
+
+func sanitizeMessagesForEstimate(messages []api.Message) []api.Message {
+	requestMessages := sanitizeMessagesForRequest(messages)
+	for i := range requestMessages {
+		// Image token accounting is model-specific. Without the active model's
+		// tokenizer and vision accounting, raw image bytes/base64 make the
+		// estimate look much larger than the prompt the model actually sees.
+		requestMessages[i].Images = nil
+	}
+	return requestMessages
+}
+
+func compactionFormatForEstimate(format string) (json.RawMessage, bool) {
+	format = strings.TrimSpace(format)
+	if format == "" {
+		return nil, false
+	}
+	if format == "json" {
+		return json.RawMessage(`"json"`), true
+	}
+	if !json.Valid([]byte(format)) {
+		return nil, false
+	}
+	return json.RawMessage(format), true
+}
+
+func compactionPrompt(previousSummary string, archive []api.Message, maxTokens int) (string, error) {
+	messages := make([]api.Message, 0, len(archive))
+	for _, msg := range archive {
+		msg.Thinking = ""
+		msg.Images = nil
+		messages = append(messages, msg)
+	}
+	return renderCompactionPrompt(previousSummary, fitCompactionMessagesToBudget(previousSummary, messages, maxTokens))
+}
+
+func renderCompactionPrompt(previousSummary string, messages []api.Message) (string, error) {
+	payload, err := json.MarshalIndent(messages, "", "  ")
+	if err != nil {
+		return "", fmt.Errorf("marshal compaction messages: %w", err)
+	}
+
+	var b strings.Builder
+	if strings.TrimSpace(previousSummary) != "" {
+		b.WriteString("Previous summary:\n")
+		b.WriteString(strings.TrimSpace(previousSummary))
+		b.WriteString("\n\n")
+	}
+	b.WriteString("Messages to archive as JSON:\n")
+	b.Write(payload)
+	return b.String(), nil
+}
+
+func fitCompactionMessagesToBudget(previousSummary string, messages []api.Message, maxTokens int) []api.Message {
+	if maxTokens <= 0 {
+		return messages
+	}
+	fitted := append([]api.Message(nil), messages...)
+	for range 16 {
+		body, err := renderCompactionPrompt(previousSummary, fitted)
+		if err != nil || estimateCompactionTokens(body) <= maxTokens {
+			return fitted
+		}
+
+		idx := largestCompactionContentMessage(fitted)
+		if idx < 0 {
+			return fitted
+		}
+		overageTokens := estimateCompactionTokens(body) - maxTokens
+		currentRunes := len([]rune(fitted[idx].Content))
+		nextRunes := currentRunes - overageTokens*4 - 256
+		if nextRunes >= currentRunes {
+			nextRunes = currentRunes / 2
+		}
+		fitted[idx].Content = truncateToolResultContentTo(fitted[idx].Content, nextRunes)
+	}
+	return fitted
+}
+
+func largestCompactionContentMessage(messages []api.Message) int {
+	idx := -1
+	size := 0
+	for i, msg := range messages {
+		n := len([]rune(msg.Content))
+		if n > size {
+			idx = i
+			size = n
+		}
+	}
+	return idx
+}
+
+func splitCompactionMessages(messages []api.Message, keepUserTurns int) (prefix []api.Message, previousSummary string, archive []api.Message, suffix []api.Message, keptUserTurns int, ok bool) {
+	if keepUserTurns < 0 {
+		keepUserTurns = defaultCompactionKeepUserTurns
+	}
+
+	start := 0
+	for start < len(messages) && messages[start].Role == "system" && !isCompactionSummary(messages[start]) {
+		prefix = append(prefix, messages[start])
+		start++
+	}
+
+	candidates := make([]api.Message, 0, len(messages)-start)
+	for i := start; i < len(messages); i++ {
+		msg := messages[i]
+		if isCompactionSummary(msg) {
+			previousSummary = CompactionSummaryText(msg.Content)
+			continue
+		}
+		if isCompactionToolCall(msg) {
+			if i+1 < len(messages) && isCompactionSummary(messages[i+1]) {
+				previousSummary = CompactionSummaryText(messages[i+1].Content)
+				i++
+			}
+			continue
+		}
+		candidates = append(candidates, msg)
+	}
+
+	userTurnIndexes := make([]int, 0, keepUserTurns)
+	for i := len(candidates) - 1; i >= 0; i-- {
+		if candidates[i].Role == "user" {
+			userTurnIndexes = append(userTurnIndexes, i)
+		}
+	}
+	keptUserTurns = keepUserTurns
+	if len(userTurnIndexes) <= keptUserTurns {
+		keptUserTurns = len(userTurnIndexes) - 1
+	}
+	if keptUserTurns < 0 {
+		keptUserTurns = 0
+	}
+
+	suffixStart := len(candidates)
+	if keptUserTurns > 0 {
+		suffixStart = userTurnIndexes[keptUserTurns-1]
+	}
+	if suffixStart <= 0 || len(candidates[:suffixStart]) == 0 {
+		return prefix, previousSummary, nil, nil, keptUserTurns, false
+	}
+
+	return prefix, previousSummary, candidates[:suffixStart], candidates[suffixStart:], keptUserTurns, true
+}
+
+func isCompactionToolName(name string) bool {
+	return name == CompactionToolName
+}
+
+func isCompactionSummary(msg api.Message) bool {
+	return (msg.Role == "user" || msg.Role == "system" || (msg.Role == "tool" && isCompactionToolName(msg.ToolName))) &&
+		strings.HasPrefix(msg.Content, CompactionSummaryMessagePrefix)
+}
+
+// IsCompactionSummary reports whether msg uses the canonical compaction
+// summary message representation.
+func IsCompactionSummary(msg api.Message) bool {
+	return isCompactionSummary(msg)
+}
+
+// CompactionSummaryContent returns the user-visible summary from msg when it
+// is a canonical compaction summary.
+func CompactionSummaryContent(msg api.Message) (string, bool) {
+	if !isCompactionSummary(msg) {
+		return "", false
+	}
+	return CompactionSummaryText(msg.Content), true
+}
+
+// IsCompactionToolResult reports whether msg is the synthetic tool result used
+// to represent compaction in message history.
+func IsCompactionToolResult(msg api.Message) bool {
+	return msg.Role == "tool" && (isCompactionToolName(msg.ToolName) || msg.ToolCallID == CompactionToolCallID)
+}
+
+// IsCompactionToolCall reports whether msg is the synthetic assistant tool
+// call paired with a compaction summary result.
+func IsCompactionToolCall(msg api.Message) bool {
+	return isCompactionToolCall(msg)
+}
+
+func isCompactionToolCall(msg api.Message) bool {
+	if msg.Role != "assistant" {
+		return false
+	}
+	for _, call := range msg.ToolCalls {
+		if isCompactionToolName(call.Function.Name) {
+			return true
+		}
+	}
+	return false
+}
+
+// CompactionSummaryText reverses CompactionSummaryMessages, returning the
+// user-visible summary text with the prefix and any continuation instruction
+// removed.
+func CompactionSummaryText(content string) string {
+	return strings.TrimSpace(strings.TrimSuffix(
+		strings.TrimSpace(strings.TrimPrefix(content, CompactionSummaryMessagePrefix)),
+		CompactionContinueInstruction,
+	))
+}
+
+func intOption(options map[string]any, key string) int {
+	if options == nil {
+		return 0
+	}
+	switch v := options[key].(type) {
+	case int:
+		return v
+	case int64:
+		return int(v)
+	case float64:
+		return int(v)
+	case float32:
+		return int(v)
+	case json.Number:
+		n, _ := v.Int64()
+		return int(n)
+	default:
+		return 0
+	}
+}
@@ -0,0 +1,773 @@
+package agent
+
+import (
+	"context"
+	"net/http"
+	"strings"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+)
+
+type scriptedCompactionClient struct {
+	responses [][]api.ChatResponse
+	errs      []error
+	requests  []*api.ChatRequest
+}
+
+func (c *scriptedCompactionClient) Chat(_ context.Context, req *api.ChatRequest, fn api.ChatResponseFunc) error {
+	c.requests = append(c.requests, req)
+	i := len(c.requests) - 1
+	if i < len(c.responses) {
+		for _, response := range c.responses[i] {
+			if err := fn(response); err != nil {
+				return err
+			}
+		}
+	}
+	if i < len(c.errs) {
+		return c.errs[i]
+	}
+	return nil
+}
+
+func assertCompactionSummaryPair(t *testing.T, messages []api.Message) {
+	t.Helper()
+	if len(messages) != 2 {
+		t.Fatalf("compaction summary pair len = %d, want 2: %#v", len(messages), messages)
+	}
+	if messages[0].Role != "assistant" || len(messages[0].ToolCalls) != 1 || messages[0].ToolCalls[0].Function.Name != CompactionToolName {
+		t.Fatalf("compaction assistant message = %#v", messages[0])
+	}
+	if messages[0].ToolCalls[0].Function.Arguments.Len() != 0 {
+		t.Fatalf("compaction summary tool call should not have arguments: %#v", messages[0].ToolCalls[0].Function.Arguments.ToMap())
+	}
+	if messages[1].Role != "tool" || messages[1].ToolName != CompactionToolName || messages[1].ToolCallID != messages[0].ToolCalls[0].ID {
+		t.Fatalf("compaction tool result = %#v", messages[1])
+	}
+	if !strings.HasPrefix(messages[1].Content, CompactionSummaryMessagePrefix) {
+		t.Fatalf("compaction tool result missing summary prefix: %#v", messages[1])
+	}
+}
+
+func TestSimpleCompactorSummarizesOldMessages(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "summary"}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 16000,
+		KeepUserTurns:       2,
+		Threshold:           0.5,
+	}}
+
+	messages := []api.Message{
+		{Role: "system", Content: "stay pinned"},
+		{Role: "user", Content: "old request"},
+		{Role: "assistant", Content: "old answer", Thinking: "hidden"},
+		{Role: "user", Content: "recent one"},
+		{Role: "assistant", Content: "recent answer"},
+		{Role: "user", Content: "recent two"},
+	}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		ChatID:   "chat-1",
+		Model:    "model",
+		Messages: messages,
+		Latest:   api.ChatResponse{Metrics: api.Metrics{PromptEvalCount: 12000}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted {
+		t.Fatal("expected compaction")
+	}
+	compacted := result.Messages
+	if len(compacted) != 6 {
+		t.Fatalf("compacted messages = %d, want 6", len(compacted))
+	}
+	if compacted[0].Content != "stay pinned" {
+		t.Fatalf("first message = %#v", compacted[0])
+	}
+	if result.Summary != "summary" {
+		t.Fatalf("result summary = %q", result.Summary)
+	}
+	assertCompactionSummaryPair(t, compacted[1:3])
+	if compacted[3].Content != "recent one" || compacted[5].Content != "recent two" {
+		t.Fatalf("recent turns were not kept: %#v", compacted)
+	}
+	if len(client.requests) != 1 {
+		t.Fatalf("summary requests = %d, want 1", len(client.requests))
+	}
+	if strings.Contains(client.requests[0].Messages[1].Content, "hidden") {
+		t.Fatal("compaction prompt should omit thinking")
+	}
+}
+
+func TestSimpleCompactorKeepsOnlySummaryForSmallContext(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "small context summary"}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: compactOnlySummaryContextTokens - 1,
+		KeepUserTurns:       3,
+		Threshold:           0.5,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		ChatID:       "chat-1",
+		Model:        "model",
+		ContinueTask: true,
+		Messages: []api.Message{
+			{Role: "system", Content: "pinned"},
+			{Role: "user", Content: "old request"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "latest request"},
+		},
+		Latest: api.ChatResponse{Metrics: api.Metrics{PromptEvalCount: 12000}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted {
+		t.Fatal("expected compaction")
+	}
+	if len(result.Messages) != 3 {
+		t.Fatalf("messages = %#v, want system plus compaction summary pair", result.Messages)
+	}
+	if result.Messages[0].Content != "pinned" {
+		t.Fatalf("leading system message not kept: %#v", result.Messages)
+	}
+	assertCompactionSummaryPair(t, result.Messages[1:])
+	if !strings.Contains(result.Messages[2].Content, CompactionContinueInstruction) {
+		t.Fatalf("tool result missing continue instruction: %q", result.Messages[2].Content)
+	}
+}
+
+func TestSimpleCompactorAddsContinueTaskInstructionOnlyToToolResult(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "summary"}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 100,
+		KeepUserTurns:       1,
+		Threshold:           0.5,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		ChatID:       "chat-1",
+		Model:        "model",
+		ContinueTask: true,
+		Messages: []api.Message{
+			{Role: "user", Content: "old request"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "recent request"},
+		},
+		Latest: api.ChatResponse{Metrics: api.Metrics{PromptEvalCount: 75}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Summary != "summary" {
+		t.Fatalf("result summary = %q", result.Summary)
+	}
+	content := result.Messages[1].Content
+	if !strings.Contains(content, CompactionContinueInstruction) {
+		t.Fatalf("tool result missing continue instruction: %q", content)
+	}
+	if got := CompactionSummaryText(content); got != "summary" {
+		t.Fatalf("visible summary text = %q", got)
+	}
+}
+
+func TestSimpleCompactorTruncatesOversizedSummary(t *testing.T) {
+	longSummary := strings.Repeat("x", maxCompactionSummaryBytes+1024)
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: longSummary}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 100,
+		KeepUserTurns:       1,
+		Threshold:           0.5,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		ChatID: "chat-1",
+		Model:  "model",
+		Messages: []api.Message{
+			{Role: "user", Content: "old one"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "recent one"},
+		},
+		Latest: api.ChatResponse{Metrics: api.Metrics{PromptEvalCount: 75}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted {
+		t.Fatal("expected compaction")
+	}
+	if len(result.Summary) > maxCompactionSummaryBytes {
+		t.Fatalf("summary bytes = %d, want <= %d", len(result.Summary), maxCompactionSummaryBytes)
+	}
+	if !strings.HasSuffix(result.Summary, compactionSummaryTruncated) {
+		t.Fatalf("summary missing truncation marker")
+	}
+	if !strings.Contains(result.Messages[1].Content, compactionSummaryTruncated) {
+		t.Fatalf("compacted message missing truncation marker: %#v", result.Messages)
+	}
+}
+
+func TestSimpleCompactorRetriesEmptySummaryWithThinkFalse(t *testing.T) {
+	client := &scriptedCompactionClient{
+		responses: [][]api.ChatResponse{
+			{{Message: api.Message{Role: "assistant", Thinking: "internal summary plan"}}},
+			{{Message: api.Message{Role: "assistant", Content: "fallback summary"}}},
+		},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 100,
+		KeepUserTurns:       1,
+		Threshold:           0.5,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		Model: "model",
+		Messages: []api.Message{
+			{Role: "user", Content: "old request"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "recent request"},
+		},
+		Force: true,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted || result.Summary != "fallback summary" {
+		t.Fatalf("compaction result = %#v", result)
+	}
+	if len(client.requests) != 2 {
+		t.Fatalf("summary requests = %d, want 2", len(client.requests))
+	}
+	if client.requests[0].Think != nil {
+		t.Fatalf("first summary request think = %#v, want nil", client.requests[0].Think)
+	}
+	if client.requests[1].Think == nil || client.requests[1].Think.Value != false {
+		t.Fatalf("fallback summary request think = %#v, want false", client.requests[1].Think)
+	}
+}
+
+func TestSimpleCompactorIgnoresUnsupportedThinkFalseFallback(t *testing.T) {
+	client := &scriptedCompactionClient{
+		responses: [][]api.ChatResponse{
+			{{Message: api.Message{Role: "assistant", Thinking: "internal summary plan"}}},
+			nil,
+		},
+		errs: []error{
+			nil,
+			api.StatusError{StatusCode: http.StatusBadRequest, ErrorMessage: "model does not support thinking"},
+		},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 100,
+		KeepUserTurns:       1,
+		Threshold:           0.5,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		Model: "model",
+		Messages: []api.Message{
+			{Role: "user", Content: "old request"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "recent request"},
+		},
+		Force: true,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Compacted || result.Reason != "summary was empty" {
+		t.Fatalf("compaction result = %#v", result)
+	}
+	if len(client.requests) != 2 {
+		t.Fatalf("summary requests = %d, want 2", len(client.requests))
+	}
+	if client.requests[1].Think == nil || client.requests[1].Think.Value != false {
+		t.Fatalf("fallback summary request think = %#v, want false", client.requests[1].Think)
+	}
+}
+
+func TestSimpleCompactorFallsBackToUnsetThinkWhenThinkFalseUnsupported(t *testing.T) {
+	client := &scriptedCompactionClient{
+		responses: [][]api.ChatResponse{
+			{{Message: api.Message{Role: "assistant", Thinking: "internal summary plan"}}},
+			nil,
+			{{Message: api.Message{Role: "assistant", Content: "unset think summary"}}},
+		},
+		errs: []error{
+			nil,
+			api.StatusError{StatusCode: http.StatusBadRequest, ErrorMessage: "think level is not supported"},
+			nil,
+		},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 100,
+		KeepUserTurns:       1,
+		Threshold:           0.5,
+	}}
+	thinkHigh := &api.ThinkValue{Value: "high"}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		Model: "model",
+		Messages: []api.Message{
+			{Role: "user", Content: "old request"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "recent request"},
+		},
+		Think: thinkHigh,
+		Force: true,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted || result.Summary != "unset think summary" {
+		t.Fatalf("compaction result = %#v", result)
+	}
+	if len(client.requests) != 3 {
+		t.Fatalf("summary requests = %d, want 3", len(client.requests))
+	}
+	if client.requests[0].Think != thinkHigh {
+		t.Fatalf("first summary request think = %#v, want original", client.requests[0].Think)
+	}
+	if client.requests[1].Think == nil || client.requests[1].Think.Value != false {
+		t.Fatalf("fallback summary request think = %#v, want false", client.requests[1].Think)
+	}
+	if client.requests[2].Think != nil {
+		t.Fatalf("unsupported fallback retry think = %#v, want nil", client.requests[2].Think)
+	}
+}
+
+func TestSimpleCompactorKeepsFewerTurnsForShortChats(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "short summary"}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 16000,
+		KeepUserTurns:       3,
+		Threshold:           0.5,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		ChatID: "chat-1",
+		Model:  "model",
+		Messages: []api.Message{
+			{Role: "user", Content: "old request"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "latest request"},
+		},
+		Latest: api.ChatResponse{Metrics: api.Metrics{PromptEvalCount: 12000}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted {
+		t.Fatal("expected compaction")
+	}
+	if len(result.Messages) != 3 {
+		t.Fatalf("messages = %#v, want compaction tool pair plus latest request", result.Messages)
+	}
+	assertCompactionSummaryPair(t, result.Messages[:2])
+	if result.Messages[2].Content != "latest request" {
+		t.Fatalf("latest turn was not kept: %#v", result.Messages)
+	}
+}
+
+func TestSimpleCompactorCanArchiveWholeShortChat(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "whole summary"}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 100,
+		KeepUserTurns:       3,
+		Threshold:           0.5,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		ChatID: "chat-1",
+		Model:  "model",
+		Messages: []api.Message{
+			{Role: "user", Content: "only request"},
+			{Role: "assistant", Content: "only answer"},
+		},
+		Latest: api.ChatResponse{Metrics: api.Metrics{PromptEvalCount: 75}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted {
+		t.Fatal("expected compaction")
+	}
+	if len(result.Messages) != 2 {
+		t.Fatalf("messages = %#v, want only compaction tool pair", result.Messages)
+	}
+	assertCompactionSummaryPair(t, result.Messages)
+}
+
+func TestSimpleCompactorSkipsBelowThreshold(t *testing.T) {
+	client := &fakeClient{}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 100,
+		Threshold:           0.8,
+	}}
+
+	messages := []api.Message{
+		{Role: "user", Content: "one"},
+		{Role: "user", Content: "two"},
+		{Role: "user", Content: "three"},
+		{Role: "user", Content: "four"},
+		{Role: "user", Content: "five"},
+	}
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		Model:    "model",
+		Messages: messages,
+		Latest:   api.ChatResponse{Metrics: api.Metrics{PromptEvalCount: 50}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Compacted {
+		t.Fatal("did not expect compaction")
+	}
+	if result.Due {
+		t.Fatal("below-threshold compaction should not be due")
+	}
+	if len(result.Messages) != len(messages) {
+		t.Fatalf("messages changed below threshold: %#v", result.Messages)
+	}
+	if len(client.requests) != 0 {
+		t.Fatalf("summary requests = %d, want 0", len(client.requests))
+	}
+}
+
+func TestSimpleCompactorUsesEstimatedMessagesWhenPromptEvalMissing(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "estimated summary"}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 100,
+		KeepUserTurns:       1,
+		Threshold:           0.8,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		Model: "model",
+		Messages: []api.Message{
+			{Role: "user", Content: "old request"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "read large output"},
+			{Role: "assistant", ToolCalls: []api.ToolCall{{
+				ID: "call-1",
+				Function: api.ToolCallFunction{
+					Name: "read",
+				},
+			}}},
+			{Role: "tool", ToolName: "read", ToolCallID: "call-1", Content: strings.Repeat("x", 360)},
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Due || !result.Compacted {
+		t.Fatalf("expected estimate-driven compaction, got %#v", result)
+	}
+	if result.Summary != "estimated summary" {
+		t.Fatalf("summary = %q", result.Summary)
+	}
+}
+
+func TestSimpleCompactorEstimateIncludesRequestPreamble(t *testing.T) {
+	compactor := &SimpleCompactor{Client: nil, Options: CompactionOptions{
+		ContextWindowTokens: 100,
+		Threshold:           0.8,
+	}}
+
+	if !compactor.shouldCompact(CompactionRequest{
+		SystemPrompt: strings.Repeat("system ", 360),
+		Messages:     []api.Message{{Role: "user", Content: "tiny"}},
+	}) {
+		t.Fatal("system prompt should count toward compaction estimate")
+	}
+
+	if !compactor.shouldCompact(CompactionRequest{
+		Messages: []api.Message{{Role: "user", Content: "tiny"}},
+		Tools: api.Tools{{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name:        "verbose_tool",
+				Description: strings.Repeat("description ", 360),
+			},
+		}},
+	}) {
+		t.Fatal("tool definitions should count toward compaction estimate")
+	}
+}
+
+func TestCompactionPromptFitsBudgetByTruncatingLargeToolOutput(t *testing.T) {
+	largeToolOutput := strings.Repeat("x", 10_000)
+	body, err := compactionPrompt("", []api.Message{
+		{Role: "user", Content: "what changed?"},
+		{Role: "assistant", ToolCalls: []api.ToolCall{{
+			ID: "call-1",
+			Function: api.ToolCallFunction{
+				Name: "bash",
+			},
+		}}},
+		{Role: "tool", ToolName: "bash", ToolCallID: "call-1", Content: largeToolOutput},
+	}, 300)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if estimateCompactionTokens(body) > 300 {
+		t.Fatalf("compaction prompt tokens = %d, want <= 300", estimateCompactionTokens(body))
+	}
+	if strings.Count(body, "x") >= len(largeToolOutput) {
+		t.Fatal("large tool output was not truncated")
+	}
+	if !strings.Contains(body, "[tool output truncated: showing first ~") {
+		t.Fatalf("truncation marker missing from compaction prompt: %q", body)
+	}
+}
+
+func TestCompactionPromptRetruncatesAlreadyTruncatedToolOutput(t *testing.T) {
+	alreadyTruncated := strings.Repeat("x", 7000) + "\n\n[tool output truncated: showing first ~100 tokens and last ~100 tokens; omitted ~99999 tokens. Use a narrower command, line range, or search query if more detail is needed.]\n\n" + strings.Repeat("y", 7000)
+	body, err := compactionPrompt("", []api.Message{
+		{Role: "user", Content: "what changed?"},
+		{Role: "assistant", ToolCalls: []api.ToolCall{{
+			ID: "call-1",
+			Function: api.ToolCallFunction{
+				Name: "bash",
+			},
+		}}},
+		{Role: "tool", ToolName: "bash", ToolCallID: "call-1", Content: alreadyTruncated},
+	}, 300)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if estimateCompactionTokens(body) > 300 {
+		t.Fatalf("compaction prompt tokens = %d, want <= 300", estimateCompactionTokens(body))
+	}
+	if strings.Count(body, "x")+strings.Count(body, "y") >= 14_000 {
+		t.Fatal("already-truncated tool output was not truncated again")
+	}
+	if !strings.Contains(body, "[tool output truncated: showing first ~") {
+		t.Fatalf("truncation marker missing from compaction prompt: %q", body)
+	}
+}
+
+func TestCompactionSummaryTextStripsPrefix(t *testing.T) {
+	content := compactionSummaryMessageForTask("worked on branch changes", false)
+	if got := CompactionSummaryText(content); got != "worked on branch changes" {
+		t.Fatalf("summary text = %q", got)
+	}
+}
+
+func TestCompactionSummaryCanTellModelToContinueTask(t *testing.T) {
+	content := compactionSummaryMessageForTask("worked on branch changes", true)
+	if !strings.Contains(content, CompactionContinueInstruction) {
+		t.Fatalf("summary message missing continue instruction: %q", content)
+	}
+	if got := CompactionSummaryText(content); got != "worked on branch changes" {
+		t.Fatalf("summary text = %q", got)
+	}
+}
+
+func TestResolveContextWindowTokensPrefersExplicitNumCtx(t *testing.T) {
+	tests := []struct {
+		name       string
+		options    map[string]any
+		configured int
+		want       int
+	}{
+		{
+			name:       "explicit smaller num ctx",
+			options:    map[string]any{"num_ctx": 4096},
+			configured: 8192,
+			want:       4096,
+		},
+		{
+			name:       "explicit num ctx can exceed configured metadata",
+			options:    map[string]any{"num_ctx": 131072},
+			configured: 8192,
+			want:       131072,
+		},
+		{
+			name:       "metadata without explicit num ctx",
+			configured: 32768,
+			want:       32768,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := ResolveContextWindowTokens(tt.options, tt.configured); got != tt.want {
+				t.Fatalf("ResolveContextWindowTokens() = %d, want %d", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestSimpleCompactorForceCompactsWithoutPromptEvalCount(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "forced summary"}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 100,
+		KeepUserTurns:       1,
+		Threshold:           0.8,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		Model: "model",
+		Messages: []api.Message{
+			{Role: "user", Content: "old"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "recent"},
+		},
+		Force: true,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Due || !result.Compacted {
+		t.Fatalf("forced compaction result = %#v", result)
+	}
+	if result.Summary != "forced summary" {
+		t.Fatalf("summary = %q", result.Summary)
+	}
+}
+
+func TestSimpleCompactorDefaultsToKeepingThreeUserTurns(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "summary"}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 16000,
+		Threshold:           0.5,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		ChatID: "chat-1",
+		Model:  "model",
+		Messages: []api.Message{
+			{Role: "user", Content: "old"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "one"},
+			{Role: "assistant", Content: "one answer"},
+			{Role: "user", Content: "two"},
+			{Role: "assistant", Content: "two answer"},
+			{Role: "user", Content: "three"},
+		},
+		Latest: api.ChatResponse{Metrics: api.Metrics{PromptEvalCount: 12000}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted {
+		t.Fatal("expected compaction")
+	}
+	assertCompactionSummaryPair(t, result.Messages[:2])
+	if got := result.Messages[2].Content; got != "one" {
+		t.Fatalf("first kept turn = %q, want one", got)
+	}
+}
+
+func TestSimpleCompactorCarriesPreviousSummary(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "new summary"}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 16000,
+		KeepUserTurns:       1,
+		Threshold:           0.5,
+	}}
+
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		Model: "model",
+		Messages: []api.Message{
+			{Role: "system", Content: CompactionSummaryMessagePrefix + "old summary"},
+			{Role: "user", Content: "old"},
+			{Role: "assistant", Content: "old answer"},
+			{Role: "user", Content: "recent"},
+		},
+		Latest: api.ChatResponse{Metrics: api.Metrics{PromptEvalCount: 12000}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted {
+		t.Fatal("expected compaction")
+	}
+	if !strings.Contains(client.requests[0].Messages[1].Content, "Previous summary:\nold summary") {
+		t.Fatalf("previous summary missing from request: %q", client.requests[0].Messages[1].Content)
+	}
+}
+
+func TestSimpleCompactorCarriesPreviousToolSummaryAndPlacesNewSummaryBeforeKeptSuffix(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "new summary"}},
+		}},
+	}
+	compactor := &SimpleCompactor{Client: client, Options: CompactionOptions{
+		ContextWindowTokens: 16000,
+		KeepUserTurns:       1,
+		Threshold:           0.5,
+	}}
+
+	messages := []api.Message{
+		{Role: "user", Content: "kept before old summary"},
+		CompactionSummaryMessages("old summary", false)[0],
+		CompactionSummaryMessages("old summary", false)[1],
+		{Role: "user", Content: "latest request"},
+	}
+	result, err := compactor.MaybeCompact(context.Background(), CompactionRequest{
+		Model:    "model",
+		Messages: messages,
+		Latest:   api.ChatResponse{Metrics: api.Metrics{PromptEvalCount: 12000}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted {
+		t.Fatal("expected compaction")
+	}
+	if !strings.Contains(client.requests[0].Messages[1].Content, "Previous summary:\nold summary") {
+		t.Fatalf("previous summary missing from request: %q", client.requests[0].Messages[1].Content)
+	}
+	if len(result.Messages) != 3 {
+		t.Fatalf("messages = %#v, want compaction pair plus latest request", result.Messages)
+	}
+	assertCompactionSummaryPair(t, result.Messages[:2])
+	if result.Messages[2].Content != "latest request" {
+		t.Fatalf("kept suffix = %#v", result.Messages)
+	}
+}
@@ -0,0 +1,79 @@
+package agent
+
+import (
+	"context"
+
+	"github.com/ollama/ollama/api"
+)
+
+type EventType string
+
+const (
+	EventMessageDelta       EventType = "message_delta"
+	EventThinkingDelta      EventType = "thinking_delta"
+	EventToolCallDetected   EventType = "tool_call_detected"
+	EventToolStarted        EventType = "tool_started"
+	EventToolFinished       EventType = "tool_finished"
+	EventCompactionStarted  EventType = "compaction_started"
+	EventCompactionProgress EventType = "compaction_progress"
+	EventCompacted          EventType = "compacted"
+	EventCompactionSkipped  EventType = "compaction_skipped"
+	EventRunFinished        EventType = "run_finished"
+	EventError              EventType = "error"
+)
+
+type Event struct {
+	Type       EventType      `json:"type"`
+	RunID      string         `json:"runId,omitempty"`
+	ChatID     string         `json:"chatId,omitempty"`
+	Model      string         `json:"model,omitempty"`
+	Status     string         `json:"status,omitempty"`
+	ToolCallID string         `json:"toolCallId,omitempty"`
+	ToolName   string         `json:"toolName,omitempty"`
+	WorkingDir string         `json:"workingDir,omitempty"`
+	Content    string         `json:"content,omitempty"`
+	Thinking   string         `json:"thinking,omitempty"`
+	ToolCalls  []api.ToolCall `json:"toolCalls,omitempty"`
+	Messages   []api.Message  `json:"messages,omitempty"`
+	Args       map[string]any `json:"args,omitempty"`
+	Tokens     int            `json:"tokens,omitempty"`
+	Error      string         `json:"error,omitempty"`
+}
+
+type EventSink interface {
+	Emit(Event) error
+}
+
+type EventSinkFunc func(Event) error
+
+func (fn EventSinkFunc) Emit(event Event) error {
+	if fn == nil {
+		return nil
+	}
+	return fn(event)
+}
+
+func (s *Session) emit(event Event) error {
+	if s == nil {
+		return nil
+	}
+	var firstErr error
+	for _, sink := range s.EventSinks {
+		if sink == nil {
+			continue
+		}
+		if err := sink.Emit(event); err != nil && firstErr == nil {
+			firstErr = err
+		}
+	}
+	return firstErr
+}
+
+func (s *Session) emitIgnoringCanceled(ctx context.Context, event Event) error {
+	err := s.emit(event)
+	if err != nil && ctx != nil && ctx.Err() != nil {
+		//nolint:nilerr // Event sinks may close during cancellation; cancellation is not a user-facing emit failure.
+		return nil
+	}
+	return err
+}
@@ -0,0 +1,97 @@
+package agent
+
+import (
+	"context"
+	"fmt"
+	"sort"
+
+	"github.com/ollama/ollama/api"
+)
+
+type ToolContext struct {
+	WorkingDir string
+}
+
+type ToolResult struct {
+	Content    string
+	WorkingDir string
+}
+
+type Tool interface {
+	Name() string
+	Description() string
+	Schema() api.ToolFunction
+	Execute(context.Context, ToolContext, map[string]any) (ToolResult, error)
+}
+
+type ApprovalRequired interface {
+	RequiresApproval(map[string]any) bool
+}
+
+type Registry struct {
+	tools map[string]Tool
+}
+
+func (r *Registry) Register(tool Tool) {
+	if r == nil || tool == nil {
+		return
+	}
+	if r.tools == nil {
+		r.tools = make(map[string]Tool)
+	}
+	r.tools[tool.Name()] = tool
+}
+
+func (r *Registry) Get(name string) (Tool, bool) {
+	if r == nil {
+		return nil, false
+	}
+	tool, ok := r.tools[name]
+	return tool, ok
+}
+
+func (r *Registry) Names() []string {
+	if r == nil {
+		return nil
+	}
+	names := make([]string, 0, len(r.tools))
+	for name := range r.tools {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	return names
+}
+
+func (r *Registry) Tools() api.Tools {
+	if r == nil {
+		return nil
+	}
+	names := r.Names()
+	apiTools := make(api.Tools, 0, len(names))
+	for _, name := range names {
+		tool := r.tools[name]
+		apiTools = append(apiTools, api.Tool{
+			Type:     "function",
+			Function: tool.Schema(),
+		})
+	}
+	return apiTools
+}
+
+func (r *Registry) Execute(ctx context.Context, toolCtx ToolContext, call api.ToolCall) (ToolResult, error) {
+	tool, ok := r.Get(call.Function.Name)
+	if !ok {
+		return ToolResult{}, fmt.Errorf("unknown tool: %s", call.Function.Name)
+	}
+	return tool.Execute(ctx, toolCtx, call.Function.Arguments.ToMap())
+}
+
+func ToolRequiresApproval(tool Tool, args map[string]any) bool {
+	if tool == nil {
+		return false
+	}
+	if t, ok := tool.(ApprovalRequired); ok {
+		return t.RequiresApproval(args)
+	}
+	return false
+}
@@ -0,0 +1,1023 @@
+package agent
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/google/uuid"
+
+	"github.com/ollama/ollama/api"
+)
+
+type ChatClient interface {
+	Chat(context.Context, *api.ChatRequest, api.ChatResponseFunc) error
+}
+
+type Session struct {
+	Client           ChatClient
+	EventSinks       []EventSink
+	Tools            *Registry
+	DisableTools     bool
+	ApprovalPrompter ApprovalPrompter
+	ApprovalState    *ApprovalState
+	WorkingDir       string
+	Compactor        Compactor
+}
+
+type RunOptions struct {
+	ChatID       string
+	Model        string
+	SystemPrompt string
+	Messages     []api.Message
+	NewMessages  []api.Message
+	Format       string
+	Options      map[string]any
+	Think        *api.ThinkValue
+	KeepAlive    *api.Duration
+	// MaxToolRounds limits consecutive model/tool cycles.
+	// Zero uses the default guard; negative disables the guard for tests or
+	// special callers.
+	MaxToolRounds int
+}
+
+type RunResult struct {
+	Messages   []api.Message
+	Latest     api.ChatResponse
+	WorkingDir string
+}
+
+const (
+	defaultMaxToolRounds              = 100
+	maxToolResultRunes                = 60000
+	smallContextToolResultRunes       = 6000
+	tinyContextToolResultRunes        = 3200
+	smallContextToolResultTokenWindow = 8192
+	tinyContextToolResultTokenWindow  = 4096
+	toolTruncationMarkerReserveTokens = 64
+	toolOutputFullOmissionPrefix      = "[tool output truncated: output omitted because the context is full;"
+)
+
+type toolOutputOverflow struct {
+	toolName   string
+	toolCallID string
+	content    string
+}
+
+type toolBatchResult struct {
+	messages  []api.Message
+	stop      toolExecutionStop
+	overflows []toolOutputOverflow
+}
+
+type toolExecutionStop string
+
+const (
+	toolExecutionDenied   toolExecutionStop = "denied"
+	toolExecutionCanceled toolExecutionStop = "canceled"
+)
+
+const toolExecutionDisabledMessage = "Tool execution disabled."
+
+type runPhase int
+
+const (
+	runPhaseModel runPhase = iota
+	runPhaseTools
+	runPhaseCompact
+	runPhaseDone
+)
+
+type runState struct {
+	runID string
+	opts  RunOptions
+
+	phase runPhase
+
+	messages []api.Message
+	latest   api.ChatResponse
+
+	assistant        api.Message
+	pendingToolCalls []api.ToolCall
+	canceled         bool
+
+	toolBatch *toolBatchResult
+
+	consecutiveModelErrors int
+	toolRounds             int
+	maxToolRounds          int
+	compactionSkipNotified bool
+
+	finish runFinish
+}
+
+type runFinish struct {
+	status         string
+	ignoreCanceled bool
+	err            error
+}
+
+func (st *runState) finishDone() {
+	st.finish = runFinish{status: "done"}
+	st.phase = runPhaseDone
+}
+
+func (st *runState) finishDenied() {
+	st.finish = runFinish{status: "denied"}
+	st.phase = runPhaseDone
+}
+
+func (st *runState) finishCanceled() {
+	st.finish = runFinish{status: "canceled", ignoreCanceled: true}
+	st.phase = runPhaseDone
+}
+
+func (st *runState) finishError(err error) {
+	st.finish = runFinish{err: err}
+	st.phase = runPhaseDone
+}
+
+func (s *Session) Run(ctx context.Context, opts RunOptions) (*RunResult, error) {
+	if s == nil {
+		return nil, errors.New("nil session")
+	}
+	if s.Client == nil {
+		return nil, errors.New("agent session requires a chat client")
+	}
+	if opts.Model == "" {
+		return nil, errors.New("agent session requires a model")
+	}
+	if s.ApprovalState == nil {
+		s.ApprovalState = &ApprovalState{}
+	}
+	runID := uuid.NewString()
+	messages := make([]api.Message, 0, len(opts.Messages)+len(opts.NewMessages))
+	for _, msg := range opts.Messages {
+		messages = append(messages, sanitizeMessageForRun(msg))
+	}
+	for _, msg := range opts.NewMessages {
+		msg = sanitizeMessageForRun(msg)
+		messages = append(messages, msg)
+	}
+
+	if err := s.checkPreflightPromptBudget(opts, messages); err != nil {
+		s.emit(Event{Type: EventError, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Error: err.Error()})
+		return nil, err
+	}
+
+	st := runState{
+		runID:         runID,
+		opts:          opts,
+		phase:         runPhaseModel,
+		messages:      messages,
+		maxToolRounds: resolvedMaxToolRounds(opts.MaxToolRounds),
+	}
+	for {
+		switch st.phase {
+		case runPhaseModel:
+			if err := s.runModelStep(ctx, &st); err != nil {
+				return nil, err
+			}
+		case runPhaseTools:
+			if err := s.runToolStep(ctx, &st); err != nil {
+				return nil, err
+			}
+		case runPhaseCompact:
+			if err := s.runCompactionStep(ctx, &st); err != nil {
+				return &RunResult{Messages: st.messages, Latest: st.latest, WorkingDir: s.WorkingDir}, err
+			}
+		case runPhaseDone:
+			return s.finishRun(ctx, &st)
+		}
+	}
+}
+
+func (s *Session) runModelStep(ctx context.Context, st *runState) error {
+	opts := st.opts
+
+	assistant, pendingToolCalls, canceled, err := s.chatRound(ctx, st.runID, opts, st.messages, &st.latest)
+	if err != nil {
+		var statusErr api.StatusError
+		if errors.As(err, &statusErr) && statusErr.StatusCode >= 500 && st.consecutiveModelErrors < 2 {
+			st.consecutiveModelErrors++
+			st.messages = append(st.messages, api.Message{
+				Role:    "user",
+				Content: fmt.Sprintf("Your previous response caused an error: %s\n\nPlease try again with a valid response.", statusErr.ErrorMessage),
+			})
+			return nil
+		}
+		s.emit(Event{Type: EventError, RunID: st.runID, ChatID: opts.ChatID, Model: opts.Model, Error: err.Error()})
+		return err
+	}
+	st.consecutiveModelErrors = 0
+	st.assistant = assistant
+	st.pendingToolCalls = pendingToolCalls
+	st.canceled = canceled
+
+	if !messageEmpty(assistant) {
+		st.messages = append(st.messages, assistant)
+	}
+
+	if len(pendingToolCalls) == 0 {
+		st.toolBatch = nil
+		st.phase = runPhaseCompact
+		return nil
+	}
+
+	if canceled {
+		skipped, skipErr := s.skipToolCalls(ctx, st.runID, opts, pendingToolCalls, "Tool execution skipped because the run was canceled.")
+		if skipErr != nil {
+			s.emit(Event{Type: EventError, RunID: st.runID, ChatID: opts.ChatID, Model: opts.Model, Error: skipErr.Error()})
+			return skipErr
+		}
+		st.messages = append(st.messages, skipped...)
+		st.finishCanceled()
+		return nil
+	}
+
+	if s.DisableTools {
+		batch, skipErr := s.disabledToolCalls(ctx, st.runID, opts, st.messages, pendingToolCalls)
+		if skipErr != nil {
+			s.emit(Event{Type: EventError, RunID: st.runID, ChatID: opts.ChatID, Model: opts.Model, Error: skipErr.Error()})
+			return skipErr
+		}
+		st.messages = append(st.messages, batch.messages...)
+		st.toolBatch = &batch
+		st.phase = runPhaseCompact
+		return nil
+	}
+
+	if s.Tools == nil {
+		st.finishDone()
+		return nil
+	}
+
+	if st.maxToolRounds >= 0 && st.toolRounds >= st.maxToolRounds {
+		content := fmt.Sprintf("Tool execution skipped because the max tool-round limit of %d was reached. Send another message to continue.", st.maxToolRounds)
+		toolMessages, skipErr := s.skipToolCalls(ctx, st.runID, opts, pendingToolCalls, content)
+		if skipErr != nil {
+			s.emit(Event{Type: EventError, RunID: st.runID, ChatID: opts.ChatID, Model: opts.Model, Error: skipErr.Error()})
+			return skipErr
+		}
+		st.messages = append(st.messages, toolMessages...)
+		err := fmt.Errorf("tool round limit reached after %d rounds; send another message to continue", st.maxToolRounds)
+		s.emit(Event{Type: EventError, RunID: st.runID, ChatID: opts.ChatID, Model: opts.Model, Error: err.Error()})
+		st.finishError(err)
+		return nil
+	}
+
+	st.phase = runPhaseTools
+	return nil
+}
+
+func (s *Session) runToolStep(ctx context.Context, st *runState) error {
+	batch, err := s.executeToolCalls(ctx, st.runID, st.opts, st.messages, st.pendingToolCalls)
+	if err != nil {
+		s.emit(Event{Type: EventError, RunID: st.runID, ChatID: st.opts.ChatID, Model: st.opts.Model, Error: err.Error()})
+		return err
+	}
+
+	st.messages = append(st.messages, batch.messages...)
+	st.toolBatch = &batch
+	st.phase = runPhaseCompact
+	return nil
+}
+
+func (s *Session) runCompactionStep(ctx context.Context, st *runState) error {
+	opts := st.opts
+	var err error
+	if st.toolBatch != nil && len(st.toolBatch.overflows) > 0 {
+		st.messages, st.compactionSkipNotified, err = s.compactForToolOutputOverflow(ctx, st.runID, opts, st.messages, st.latest, st.assistant, st.toolBatch.messages, st.toolBatch.overflows, st.compactionSkipNotified)
+	} else {
+		st.messages, st.compactionSkipNotified, err = s.maybeCompact(ctx, st.runID, opts, st.messages, st.latest, st.compactionSkipNotified)
+	}
+	if err != nil {
+		s.emit(Event{Type: EventError, RunID: st.runID, ChatID: opts.ChatID, Model: opts.Model, Error: err.Error()})
+		return err
+	}
+
+	if st.toolBatch == nil {
+		if st.canceled {
+			st.finishCanceled()
+		} else {
+			st.finishDone()
+		}
+		return nil
+	}
+
+	switch st.toolBatch.stop {
+	case toolExecutionDenied:
+		st.finishDenied()
+	case toolExecutionCanceled:
+		st.finishCanceled()
+	default:
+		st.toolRounds++
+		st.assistant = api.Message{}
+		st.pendingToolCalls = nil
+		st.toolBatch = nil
+		st.phase = runPhaseModel
+	}
+	return nil
+}
+
+func (s *Session) finishRun(ctx context.Context, st *runState) (*RunResult, error) {
+	if st.finish.status != "" {
+		event := Event{Type: EventRunFinished, RunID: st.runID, ChatID: st.opts.ChatID, Model: st.opts.Model, Status: st.finish.status}
+		var err error
+		if st.finish.ignoreCanceled {
+			err = s.emitIgnoringCanceled(ctx, event)
+		} else {
+			err = s.emit(event)
+		}
+		if err != nil {
+			return nil, err
+		}
+	}
+	return &RunResult{Messages: st.messages, Latest: st.latest, WorkingDir: s.WorkingDir}, st.finish.err
+}
+
+func (s *Session) chatRound(ctx context.Context, runID string, opts RunOptions, messages []api.Message, latest *api.ChatResponse) (api.Message, []api.ToolCall, bool, error) {
+	var tools api.Tools
+	if !s.DisableTools {
+		tools = s.availableTools()
+	}
+	req := buildChatRequest(opts, messages, tools)
+
+	assistant := api.Message{Role: "assistant"}
+	var pendingToolCalls []api.ToolCall
+
+	err := s.Client.Chat(ctx, &req, func(response api.ChatResponse) error {
+		if response.Message.Role != "" {
+			assistant.Role = response.Message.Role
+		}
+
+		if response.Message.Content == "" && response.Message.Thinking == "" && len(response.Message.ToolCalls) == 0 {
+			*latest = response
+			return nil
+		}
+
+		if response.Message.Thinking != "" {
+			assistant.Thinking += response.Message.Thinking
+			if err := s.emit(Event{Type: EventThinkingDelta, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Thinking: response.Message.Thinking}); err != nil {
+				return err
+			}
+		}
+
+		if response.Message.Content != "" {
+			assistant.Content += response.Message.Content
+			if err := s.emit(Event{Type: EventMessageDelta, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Content: response.Message.Content}); err != nil {
+				return err
+			}
+		}
+
+		if len(response.Message.ToolCalls) > 0 {
+			assistant.ToolCalls = append(assistant.ToolCalls, response.Message.ToolCalls...)
+			pendingToolCalls = append(pendingToolCalls, response.Message.ToolCalls...)
+			if err := s.emit(Event{Type: EventToolCallDetected, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, ToolCalls: response.Message.ToolCalls}); err != nil {
+				return err
+			}
+		}
+
+		*latest = response
+		return nil
+	})
+	if err != nil {
+		if isContextCanceledError(ctx, err) {
+			return assistant, pendingToolCalls, true, nil
+		}
+		return assistant, pendingToolCalls, false, err
+	}
+
+	return assistant, pendingToolCalls, false, nil
+}
+
+func buildChatRequest(opts RunOptions, messages []api.Message, tools api.Tools) api.ChatRequest {
+	requestMessages := sanitizeMessagesForRequest(messages)
+	if strings.TrimSpace(opts.SystemPrompt) != "" {
+		withSystem := make([]api.Message, 0, len(requestMessages)+1)
+		withSystem = append(withSystem, api.Message{Role: "system", Content: opts.SystemPrompt})
+		requestMessages = append(withSystem, requestMessages...)
+	}
+
+	format := opts.Format
+	if format == "json" {
+		format = `"` + format + `"`
+	}
+
+	req := api.ChatRequest{
+		Model:    opts.Model,
+		Messages: requestMessages,
+		Format:   json.RawMessage(format),
+		Options:  opts.Options,
+		Think:    opts.Think,
+	}
+	if opts.KeepAlive != nil {
+		req.KeepAlive = opts.KeepAlive
+	}
+	if len(tools) > 0 {
+		req.Tools = tools
+	}
+	return req
+}
+
+func (s *Session) executeToolCalls(ctx context.Context, runID string, opts RunOptions, messages []api.Message, calls []api.ToolCall) (toolBatchResult, error) {
+	batch := toolBatchResult{
+		messages: make([]api.Message, 0, len(calls)),
+	}
+	projectedMessages := append([]api.Message(nil), messages...)
+
+	type plannedToolCall struct {
+		call       api.ToolCall
+		tool       Tool
+		toolName   string
+		args       map[string]any
+		workingDir string
+	}
+	plans := make([]plannedToolCall, 0, len(calls))
+	batchWorkingDir := s.currentWorkingDir()
+	approvalReq := ApprovalRequest{WorkingDir: batchWorkingDir}
+	for _, call := range calls {
+		toolName := call.Function.Name
+		args := call.Function.Arguments.ToMap()
+		tool, ok := s.Tools.Get(toolName)
+		plans = append(plans, plannedToolCall{
+			call:       call,
+			tool:       tool,
+			toolName:   toolName,
+			args:       args,
+			workingDir: batchWorkingDir,
+		})
+		if ok && s.needsApproval(tool, toolName, args) {
+			approvalReq.AddToolCall(call.ID, toolName, args)
+		}
+	}
+
+	if len(approvalReq.Calls) > 0 {
+		approvalResult, err := s.authorizeToolCalls(ctx, approvalReq)
+		if err != nil {
+			if ctx.Err() != nil {
+				skipped, skipErr := s.skipToolCalls(ctx, runID, opts, calls, "Tool execution skipped because the run was canceled.")
+				if skipErr != nil {
+					return toolBatchResult{}, skipErr
+				}
+				batch.messages = append(batch.messages, skipped...)
+				batch.stop = toolExecutionCanceled
+				return batch, nil
+			}
+			return toolBatchResult{}, err
+		}
+		if !approvalResult.Allow {
+			content := approvalResult.Reason
+			if content == "" {
+				content = "Tool execution denied."
+			}
+			for _, plan := range plans {
+				msg := s.toolMessageForContext(plan.toolName, plan.call.ID, content, opts, projectedMessages)
+				batch.messages = append(batch.messages, msg)
+				projectedMessages = append(projectedMessages, msg)
+				deniedContent := msg.Content
+				if emitErr := s.emit(Event{Type: EventToolFinished, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Status: "denied", ToolCallID: plan.call.ID, ToolName: plan.toolName, Args: plan.args, Content: deniedContent, Error: deniedContent}); emitErr != nil {
+					return toolBatchResult{}, emitErr
+				}
+			}
+			batch.stop = toolExecutionDenied
+			return batch, nil
+		}
+	}
+
+	for i, plan := range plans {
+		call := plan.call
+		toolName := plan.toolName
+		args := plan.args
+		if ctx.Err() != nil {
+			skipped, skipErr := s.skipToolCalls(ctx, runID, opts, calls[i:], "Tool execution skipped because the run was canceled.")
+			if skipErr != nil {
+				return toolBatchResult{}, skipErr
+			}
+			batch.messages = append(batch.messages, skipped...)
+			batch.stop = toolExecutionCanceled
+			return batch, nil
+		}
+		if plan.tool == nil {
+			content := fmt.Sprintf("Error: unknown tool: %s", toolName)
+			msg := s.toolMessageForContext(toolName, call.ID, content, opts, projectedMessages)
+			batch.messages = append(batch.messages, msg)
+			projectedMessages = append(projectedMessages, msg)
+			content = msg.Content
+			if toolOutputFullyOmitted(content) {
+				batch.overflows = append(batch.overflows, toolOutputOverflow{toolName: toolName, toolCallID: call.ID, content: fmt.Sprintf("Error: unknown tool: %s", toolName)})
+			}
+			if emitErr := s.emit(Event{Type: EventToolFinished, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Status: "failed", ToolCallID: call.ID, ToolName: toolName, Args: args, Content: content, Error: fmt.Sprintf("unknown tool: %s", toolName)}); emitErr != nil {
+				return toolBatchResult{}, emitErr
+			}
+			continue
+		}
+
+		if err := s.emit(Event{Type: EventToolStarted, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Status: "running", ToolCallID: call.ID, ToolName: toolName, WorkingDir: plan.workingDir, Args: args}); err != nil {
+			return toolBatchResult{}, err
+		}
+
+		result, err := s.Tools.Execute(ctx, ToolContext{WorkingDir: plan.workingDir}, call)
+		if err != nil {
+			rawContent := fmt.Sprintf("Error: %v", err)
+			msg := s.toolMessageForContext(toolName, call.ID, rawContent, opts, projectedMessages)
+			batch.messages = append(batch.messages, msg)
+			projectedMessages = append(projectedMessages, msg)
+			content := msg.Content
+			if toolOutputFullyOmitted(content) {
+				batch.overflows = append(batch.overflows, toolOutputOverflow{toolName: toolName, toolCallID: call.ID, content: rawContent})
+			}
+			if emitErr := s.emitIgnoringCanceled(ctx, Event{Type: EventToolFinished, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Status: "failed", ToolCallID: call.ID, ToolName: toolName, Args: args, Content: content, Error: err.Error()}); emitErr != nil {
+				return toolBatchResult{}, emitErr
+			}
+			if ctx.Err() != nil {
+				skipped, skipErr := s.skipToolCalls(ctx, runID, opts, calls[i+1:], "Tool execution skipped because the run was canceled.")
+				if skipErr != nil {
+					return toolBatchResult{}, skipErr
+				}
+				batch.messages = append(batch.messages, skipped...)
+				batch.stop = toolExecutionCanceled
+				return batch, nil
+			}
+			continue
+		}
+
+		eventWorkingDir := plan.workingDir
+		if s.applyToolWorkingDir(result.WorkingDir) {
+			eventWorkingDir = s.WorkingDir
+		}
+		rawContent := result.Content
+
+		msg := s.toolMessageForContext(toolName, call.ID, rawContent, opts, projectedMessages)
+		batch.messages = append(batch.messages, msg)
+		projectedMessages = append(projectedMessages, msg)
+		content := msg.Content
+
+		if toolOutputFullyOmitted(content) {
+			batch.overflows = append(batch.overflows, toolOutputOverflow{toolName: toolName, toolCallID: call.ID, content: rawContent})
+		}
+		if err := s.emitIgnoringCanceled(ctx, Event{Type: EventToolFinished, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Status: "done", ToolCallID: call.ID, ToolName: toolName, WorkingDir: eventWorkingDir, Args: args, Content: content}); err != nil {
+			return toolBatchResult{}, err
+		}
+		if ctx.Err() != nil {
+			skipped, skipErr := s.skipToolCalls(ctx, runID, opts, calls[i+1:], "Tool execution skipped because the run was canceled.")
+			if skipErr != nil {
+				return toolBatchResult{}, skipErr
+			}
+			batch.messages = append(batch.messages, skipped...)
+			batch.stop = toolExecutionCanceled
+			return batch, nil
+		}
+	}
+	return batch, nil
+}
+
+func (s *Session) disabledToolCalls(ctx context.Context, runID string, opts RunOptions, messages []api.Message, calls []api.ToolCall) (toolBatchResult, error) {
+	batch := toolBatchResult{
+		messages: make([]api.Message, 0, len(calls)),
+	}
+	projectedMessages := append([]api.Message(nil), messages...)
+	for _, call := range calls {
+		toolName := call.Function.Name
+		args := call.Function.Arguments.ToMap()
+		msg := s.toolMessageForContext(toolName, call.ID, toolExecutionDisabledMessage, opts, projectedMessages)
+		batch.messages = append(batch.messages, msg)
+		projectedMessages = append(projectedMessages, msg)
+		if emitErr := s.emitIgnoringCanceled(ctx, Event{Type: EventToolFinished, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Status: "disabled", ToolCallID: call.ID, ToolName: toolName, Args: args, Content: msg.Content, Error: msg.Content}); emitErr != nil {
+			return toolBatchResult{}, emitErr
+		}
+	}
+	return batch, nil
+}
+
+func (s *Session) skipToolCalls(ctx context.Context, runID string, opts RunOptions, calls []api.ToolCall, content string) ([]api.Message, error) {
+	toolMessages := make([]api.Message, 0, len(calls))
+	for _, call := range calls {
+		toolName := call.Function.Name
+		args := call.Function.Arguments.ToMap()
+		msg := toolMessage(toolName, call.ID, content)
+		toolMessages = append(toolMessages, msg)
+		if emitErr := s.emitIgnoringCanceled(ctx, Event{Type: EventToolFinished, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Status: "skipped", ToolCallID: call.ID, ToolName: toolName, Args: args, Content: msg.Content, Error: msg.Content}); emitErr != nil {
+			return nil, emitErr
+		}
+	}
+	return toolMessages, nil
+}
+
+func (s *Session) currentWorkingDir() string {
+	if s.WorkingDir != "" {
+		return s.WorkingDir
+	}
+	wd, err := os.Getwd()
+	if err != nil {
+		return ""
+	}
+	s.WorkingDir = wd
+	return s.WorkingDir
+}
+
+func (s *Session) applyToolWorkingDir(next string) bool {
+	next = strings.TrimSpace(next)
+	if next == "" {
+		return false
+	}
+	current := s.currentWorkingDir()
+	nextAbs, err := canonicalSessionPath(next)
+	if err != nil {
+		return false
+	}
+	if current == nextAbs {
+		return false
+	}
+	s.WorkingDir = nextAbs
+	return true
+}
+
+func canonicalSessionPath(path string) (string, error) {
+	abs, err := filepath.Abs(path)
+	if err != nil {
+		return "", err
+	}
+	resolved, err := filepath.EvalSymlinks(abs)
+	if err == nil {
+		return resolved, nil
+	}
+	return abs, nil
+}
+
+func isContextCanceledError(ctx context.Context, err error) bool {
+	if err == nil {
+		return false
+	}
+	if errors.Is(err, context.Canceled) {
+		return true
+	}
+	return ctx != nil && errors.Is(ctx.Err(), context.Canceled) && strings.Contains(err.Error(), "context canceled")
+}
+
+func (s *Session) maybeCompact(ctx context.Context, runID string, opts RunOptions, messages []api.Message, latest api.ChatResponse, skipNotified bool) ([]api.Message, bool, error) {
+	if s.Compactor == nil {
+		return messages, skipNotified, nil
+	}
+	req := s.compactionRequest(runID, opts, messages, latest)
+	trigger := s.autoCompactionTrigger(req)
+	if trigger != "" {
+		s.emitCompactionStarted(runID, opts, trigger)
+	}
+	result, err := s.Compactor.MaybeCompact(ctx, req)
+	if err != nil {
+		if result.Due && !skipNotified {
+			if trigger == "" {
+				trigger = "error"
+			}
+			s.emitCompactionSkipped(runID, opts, trigger, result.Reason)
+			skipNotified = true
+		}
+		return messages, skipNotified, nil
+	}
+	if !result.Compacted {
+		if result.Due && !skipNotified {
+			if trigger == "" {
+				trigger = "due"
+			}
+			s.emitCompactionSkipped(runID, opts, trigger, result.Reason)
+			skipNotified = true
+		}
+		return messages, skipNotified, nil
+	}
+	s.emitCompacted(runID, opts, result.Messages, trigger, result.Summary)
+	if err := s.checkPostCompactionPromptBudget(opts, result.Messages); err != nil {
+		return result.Messages, skipNotified, err
+	}
+	return result.Messages, skipNotified, nil
+}
+
+func (s *Session) compactForToolOutputOverflow(ctx context.Context, runID string, opts RunOptions, messages []api.Message, latest api.ChatResponse, assistant api.Message, toolMessages []api.Message, overflows []toolOutputOverflow, skipNotified bool) ([]api.Message, bool, error) {
+	if s.Compactor == nil {
+		return messages, skipNotified, nil
+	}
+
+	keepUserTurns := 0
+	req := s.compactionRequest(runID, opts, messages, latest)
+	req.Force = true
+	req.KeepUserTurns = &keepUserTurns
+	s.emitCompactionStarted(runID, opts, "tool_output")
+
+	result, err := s.Compactor.MaybeCompact(ctx, req)
+	if err != nil {
+		if result.Due && !skipNotified {
+			s.emitCompactionSkipped(runID, opts, "tool_output", result.Reason)
+			skipNotified = true
+		}
+		return messages, skipNotified, nil
+	}
+	if !result.Compacted {
+		if result.Due && !skipNotified {
+			s.emitCompactionSkipped(runID, opts, "tool_output", result.Reason)
+			skipNotified = true
+		}
+		return messages, skipNotified, nil
+	}
+
+	overflowByID := make(map[string]toolOutputOverflow, len(overflows))
+	for _, overflow := range overflows {
+		overflowByID[overflow.toolCallID] = overflow
+	}
+
+	compacted := append([]api.Message(nil), result.Messages...)
+	if !messageEmpty(assistant) {
+		compacted = append(compacted, assistant)
+	}
+
+	for _, msg := range toolMessages {
+		content := msg.Content
+		toolName := msg.ToolName
+		if overflow, ok := overflowByID[msg.ToolCallID]; ok {
+			content = overflow.content
+			if overflow.toolName != "" {
+				toolName = overflow.toolName
+			}
+		}
+		refit := s.toolMessageForPostCompactionContext(toolName, msg.ToolCallID, content, opts, compacted)
+		compacted = append(compacted, refit)
+	}
+
+	s.emitCompacted(runID, opts, compacted, "tool_output", result.Summary)
+	if err := s.checkPostCompactionPromptBudget(opts, compacted); err != nil {
+		return compacted, skipNotified, err
+	}
+	return compacted, skipNotified, nil
+}
+
+func (s *Session) compactionRequest(runID string, opts RunOptions, messages []api.Message, latest api.ChatResponse) CompactionRequest {
+	return CompactionRequest{
+		ChatID:       opts.ChatID,
+		Model:        opts.Model,
+		SystemPrompt: opts.SystemPrompt,
+		Messages:     messages,
+		Tools:        s.availableTools(),
+		Format:       opts.Format,
+		Latest:       latest,
+		Options:      opts.Options,
+		KeepAlive:    opts.KeepAlive,
+		Think:        opts.Think,
+		ContinueTask: true,
+		Progress: func(progress CompactionProgress) {
+			_ = s.emit(Event{Type: EventCompactionProgress, RunID: runID, ChatID: opts.ChatID, Model: opts.Model, Tokens: progress.Tokens})
+		},
+	}
+}
+
+func (s *Session) emitCompactionStarted(runID string, opts RunOptions, status string) {
+	_ = s.emit(Event{
+		Type:   EventCompactionStarted,
+		RunID:  runID,
+		ChatID: opts.ChatID,
+		Model:  opts.Model,
+		Status: status,
+	})
+}
+
+func (s *Session) emitCompactionSkipped(runID string, opts RunOptions, status, reason string) {
+	_ = s.emit(Event{
+		Type:    EventCompactionSkipped,
+		RunID:   runID,
+		ChatID:  opts.ChatID,
+		Model:   opts.Model,
+		Status:  status,
+		Content: CompactionSkippedMessage(reason),
+	})
+}
+
+func (s *Session) emitCompacted(runID string, opts RunOptions, messages []api.Message, status, summary string) {
+	_ = s.emit(Event{
+		Type:     EventCompacted,
+		RunID:    runID,
+		ChatID:   opts.ChatID,
+		Model:    opts.Model,
+		Status:   status,
+		Content:  summary,
+		Messages: messages,
+	})
+}
+
+func (s *Session) autoCompactionTrigger(req CompactionRequest) string {
+	if compactor, ok := s.Compactor.(*SimpleCompactor); ok && compactor != nil {
+		if req.Force {
+			return "force"
+		}
+		contextWindow := compactor.contextWindowTokens(req.Options)
+		threshold := int(float64(contextWindow) * compactor.threshold())
+		if threshold <= 0 {
+			return ""
+		}
+		if req.Latest.PromptEvalCount > 0 && req.Latest.PromptEvalCount >= threshold {
+			return "prompt_eval"
+		}
+		if estimateCompactionRequestTokens(req) >= threshold {
+			return "estimate"
+		}
+	}
+	return ""
+}
+
+func CompactionSkippedMessage(reason string) string {
+	reason = strings.TrimSpace(reason)
+	if reason == "" {
+		reason = "compaction could not run"
+	}
+	return reason
+}
+
+func resolvedMaxToolRounds(value int) int {
+	if value == 0 {
+		return defaultMaxToolRounds
+	}
+	return value
+}
+
+func (s *Session) toolMessageForContext(toolName, toolCallID, content string, opts RunOptions, messages []api.Message) api.Message {
+	maxRunes := maxToolResultRunes
+	if limit := smallContextToolResultLimitRunes(s.contextWindowTokens(opts)); limit > 0 {
+		maxRunes = min(maxRunes, limit)
+	}
+
+	msg := toolMessageWithLimit(toolName, toolCallID, content, maxRunes)
+	threshold := s.compactionThresholdTokens(opts)
+	if threshold <= 0 {
+		return msg
+	}
+
+	projected := append(append([]api.Message(nil), messages...), msg)
+	projectedTokens := s.estimateRunPromptTokens(opts, projected)
+	if projectedTokens < threshold {
+		return msg
+	}
+
+	baseTokens := s.estimateRunPromptTokens(opts, messages)
+	overheadTokens := estimateMessagesTokens([]api.Message{{
+		Role:       "tool",
+		ToolName:   toolName,
+		ToolCallID: toolCallID,
+	}})
+	// Keep oversized tool output below the compaction threshold before it is
+	// appended to history. This is especially important for <=8k contexts: the
+	// next step must still have enough room to compact and continue the same
+	// user request instead of asking the user to prompt again.
+	availableRunes := (threshold - baseTokens - overheadTokens - toolTruncationMarkerReserveTokens) * 4
+	maxRunes = min(maxRunes, max(0, availableRunes))
+	msg.Content = truncateToolResultContentTo(content, maxRunes)
+	return msg
+}
+
+func (s *Session) toolMessageForPostCompactionContext(toolName, toolCallID, content string, opts RunOptions, messages []api.Message) api.Message {
+	maxRunes := maxToolResultRunes
+	if limit := smallContextToolResultLimitRunes(s.contextWindowTokens(opts)); limit > 0 {
+		maxRunes = min(maxRunes, limit)
+	}
+
+	contextWindow := s.contextWindowTokens(opts)
+	if contextWindow <= 0 {
+		return toolMessageWithLimit(toolName, toolCallID, content, maxRunes)
+	}
+
+	baseTokens := s.estimateRunPromptTokens(opts, messages)
+	overheadTokens := estimateMessagesTokens([]api.Message{{
+		Role:       "tool",
+		ToolName:   toolName,
+		ToolCallID: toolCallID,
+	}})
+	availableRunes := (contextWindow - baseTokens - overheadTokens - toolTruncationMarkerReserveTokens) * 4
+	maxRunes = min(maxRunes, max(0, availableRunes))
+	return toolMessageWithLimit(toolName, toolCallID, content, maxRunes)
+}
+
+func toolMessageWithLimit(toolName, toolCallID, content string, maxRunes int) api.Message {
+	return api.Message{
+		Role:       "tool",
+		Content:    truncateToolResultContentTo(content, maxRunes),
+		ToolName:   toolName,
+		ToolCallID: toolCallID,
+	}
+}
+
+func smallContextToolResultLimitRunes(contextWindow int) int {
+	switch {
+	case contextWindow > 0 && contextWindow <= tinyContextToolResultTokenWindow:
+		return tinyContextToolResultRunes
+	case contextWindow > 0 && contextWindow <= smallContextToolResultTokenWindow:
+		return smallContextToolResultRunes
+	default:
+		return 0
+	}
+}
+
+func (s *Session) availableTools() api.Tools {
+	if s == nil || s.Tools == nil {
+		return nil
+	}
+	return s.Tools.Tools()
+}
+
+func (s *Session) compactionThresholdTokens(opts RunOptions) int {
+	contextWindow := s.contextWindowTokens(opts)
+	if contextWindow <= 0 {
+		return 0
+	}
+
+	configuredThreshold := 0.0
+	if compactor, ok := s.Compactor.(*SimpleCompactor); ok && compactor != nil {
+		configuredThreshold = compactor.Options.Threshold
+	}
+
+	threshold := int(float64(contextWindow) * ResolveCompactionThreshold(configuredThreshold))
+	if threshold <= 0 {
+		return 0
+	}
+	return threshold
+}
+
+func (s *Session) contextWindowTokens(opts RunOptions) int {
+	if s.Compactor == nil {
+		return 0
+	}
+
+	configuredWindow := 0
+	if compactor, ok := s.Compactor.(*SimpleCompactor); ok && compactor != nil {
+		configuredWindow = compactor.Options.ContextWindowTokens
+	}
+
+	return ResolveContextWindowTokens(opts.Options, configuredWindow)
+}
+
+func toolMessage(toolName, toolCallID, content string) api.Message {
+	return toolMessageWithLimit(toolName, toolCallID, content, maxToolResultRunes)
+}
+
+func sanitizeMessageForRun(msg api.Message) api.Message {
+	if msg.Role == "tool" {
+		msg.Content = truncateToolResultContent(msg.Content)
+	}
+	return msg
+}
+
+func sanitizeMessagesForRequest(messages []api.Message) []api.Message {
+	if len(messages) == 0 {
+		return nil
+	}
+	sanitized := make([]api.Message, len(messages))
+	for i, msg := range messages {
+		sanitized[i] = sanitizeMessageForRequest(msg)
+	}
+	return sanitized
+}
+
+func sanitizeMessageForRequest(msg api.Message) api.Message {
+	return sanitizeMessageForRun(msg)
+}
+
+func truncateToolResultContent(content string) string {
+	return truncateToolResultContentTo(content, maxToolResultRunes)
+}
+
+func truncateToolResultContentTo(content string, maxRunes int) string {
+	if maxRunes <= 0 {
+		return fmt.Sprintf("%s omitted ~%d tokens. Use a narrower command, line range, or search query if more detail is needed.]", toolOutputFullOmissionPrefix, approximateTokensFromRunes(len([]rune(content))))
+	}
+	if len(content) <= maxRunes {
+		return content
+	}
+	runes := []rune(content)
+	if len(runes) <= maxRunes {
+		return content
+	}
+	head := maxRunes * 3 / 4
+	tail := maxRunes - head
+	omitted := len(runes) - head - tail
+	marker := fmt.Sprintf(
+		"\n\n[tool output truncated: showing first ~%d tokens and last ~%d tokens; omitted ~%d tokens. Use a narrower command, line range, or search query if more detail is needed.]\n\n",
+		approximateTokensFromRunes(head),
+		approximateTokensFromRunes(tail),
+		approximateTokensFromRunes(omitted),
+	)
+	return string(runes[:head]) + marker + string(runes[len(runes)-tail:])
+}
+
+func toolOutputFullyOmitted(content string) bool {
+	return strings.HasPrefix(content, toolOutputFullOmissionPrefix)
+}
+
+func approximateTokensFromRunes(n int) int {
+	if n <= 0 {
+		return 0
+	}
+	return max(1, (n+3)/4)
+}
+
+func messageEmpty(msg api.Message) bool {
+	return msg.Content == "" && msg.Thinking == "" && len(msg.ToolCalls) == 0
+}
@@ -0,0 +1,2188 @@
+package agent
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+)
+
+type fakeClient struct {
+	calls     int
+	responses [][]api.ChatResponse
+	requests  []*api.ChatRequest
+	err       error
+}
+
+func (c *fakeClient) Chat(ctx context.Context, req *api.ChatRequest, fn api.ChatResponseFunc) error {
+	c.requests = append(c.requests, req)
+	if c.calls >= len(c.responses) {
+		return nil
+	}
+	responses := c.responses[c.calls]
+	c.calls++
+	for _, response := range responses {
+		if err := fn(response); err != nil {
+			return err
+		}
+	}
+	return c.err
+}
+
+type staticTool struct{}
+
+type approvalTestTool struct {
+	called *bool
+}
+
+type namedApprovalTestTool struct {
+	name string
+}
+
+type cwdTestTool struct{}
+
+type largeTool struct{}
+
+type preTruncatedTool struct{}
+
+type cancelingTool struct {
+	cancel context.CancelFunc
+}
+
+type cancelAfterToolCallClient struct {
+	cancel context.CancelFunc
+}
+
+type recordingCompactor struct {
+	requests []CompactionRequest
+}
+
+type oversizedCompactor struct {
+	requests []CompactionRequest
+}
+
+type recordingEventSink struct {
+	events []Event
+}
+
+func (s *recordingEventSink) Emit(event Event) error {
+	s.events = append(s.events, event)
+	return nil
+}
+
+func hasEventType(events []Event, eventType EventType) bool {
+	for _, event := range events {
+		if event.Type == eventType {
+			return true
+		}
+	}
+	return false
+}
+
+func hasEventWithTokens(events []Event, eventType EventType, tokens int) bool {
+	for _, event := range events {
+		if event.Type == eventType && event.Tokens == tokens {
+			return true
+		}
+	}
+	return false
+}
+
+func TestSessionEmitsToAllSinksAfterError(t *testing.T) {
+	errSink := EventSinkFunc(func(Event) error {
+		return errors.New("sink failed")
+	})
+	events := &recordingEventSink{}
+	session := &Session{EventSinks: []EventSink{errSink, events}}
+
+	err := session.emit(Event{Type: EventRunFinished})
+	if err == nil {
+		t.Fatal("emit should return the first sink error")
+	}
+	if !hasEventType(events.events, EventRunFinished) {
+		t.Fatalf("later sink did not receive event after earlier error: %#v", events.events)
+	}
+}
+
+func (c cancelAfterToolCallClient) Chat(ctx context.Context, req *api.ChatRequest, fn api.ChatResponseFunc) error {
+	args := api.NewToolCallFunctionArguments()
+	args.Set("value", "skip me")
+	if err := fn(api.ChatResponse{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+		ID: "call-1",
+		Function: api.ToolCallFunction{
+			Name:      "echo_tool",
+			Arguments: args,
+		},
+	}}}}); err != nil {
+		return err
+	}
+	c.cancel()
+	return context.Canceled
+}
+
+func (c *recordingCompactor) MaybeCompact(_ context.Context, req CompactionRequest) (CompactionResult, error) {
+	c.requests = append(c.requests, req)
+	result := CompactionResult{Messages: req.Messages, Due: true}
+	if len(req.Messages) > 0 && req.Messages[len(req.Messages)-1].Role == "tool" {
+		result.Messages = CompactionSummaryMessages("tool result summarized", false)
+		result.Compacted = true
+		result.Summary = "tool result summarized"
+	}
+	return result, nil
+}
+
+func (c *oversizedCompactor) MaybeCompact(_ context.Context, req CompactionRequest) (CompactionResult, error) {
+	c.requests = append(c.requests, req)
+	summary := strings.Repeat("oversized summary ", 300)
+	return CompactionResult{
+		Messages:  CompactionSummaryMessages(summary, req.ContinueTask),
+		Compacted: true,
+		Due:       true,
+		Summary:   summary,
+	}, nil
+}
+
+type recordingApprovalPrompter struct {
+	requests []ApprovalRequest
+	results  []Approval
+}
+
+func approvalStateForTest(allowAll bool, scopes map[string]bool) *ApprovalState {
+	state := &ApprovalState{}
+	state.Set(allowAll, scopes)
+	return state
+}
+
+func (staticTool) Name() string {
+	return "echo_tool"
+}
+
+func (staticTool) Description() string {
+	return "echoes a value"
+}
+
+func (staticTool) Schema() api.ToolFunction {
+	props := api.NewToolPropertiesMap()
+	props.Set("value", api.ToolProperty{Type: api.PropertyType{"string"}})
+	return api.ToolFunction{
+		Name:        "echo_tool",
+		Description: "echoes a value",
+		Parameters: api.ToolFunctionParameters{
+			Type:       "object",
+			Properties: props,
+		},
+	}
+}
+
+func (staticTool) Execute(context.Context, ToolContext, map[string]any) (ToolResult, error) {
+	return ToolResult{Content: "tool says hello"}, nil
+}
+
+func (largeTool) Name() string {
+	return "large_tool"
+}
+
+func (largeTool) Description() string {
+	return "returns a large result"
+}
+
+func (largeTool) Schema() api.ToolFunction {
+	return api.ToolFunction{
+		Name:        "large_tool",
+		Description: "returns a large result",
+		Parameters: api.ToolFunctionParameters{
+			Type: "object",
+		},
+	}
+}
+
+func (largeTool) Execute(context.Context, ToolContext, map[string]any) (ToolResult, error) {
+	return ToolResult{Content: strings.Repeat("x", maxToolResultRunes+100)}, nil
+}
+
+func (preTruncatedTool) Name() string {
+	return "pre_truncated_tool"
+}
+
+func (preTruncatedTool) Description() string {
+	return "returns a large result that is already marked as truncated"
+}
+
+func (preTruncatedTool) Schema() api.ToolFunction {
+	return api.ToolFunction{
+		Name:        "pre_truncated_tool",
+		Description: "returns a large result that is already marked as truncated",
+		Parameters: api.ToolFunctionParameters{
+			Type: "object",
+		},
+	}
+}
+
+func (preTruncatedTool) Execute(context.Context, ToolContext, map[string]any) (ToolResult, error) {
+	content := strings.Repeat("x", smallContextToolResultRunes) +
+		"\n\n[tool output truncated: showing first ~1500 tokens; omitted ~999 tokens. Use a narrower command, line range, or search query if more detail is needed.]\n\n" +
+		strings.Repeat("y", smallContextToolResultRunes)
+	return ToolResult{Content: content}, nil
+}
+
+func (t cancelingTool) Name() string {
+	return "cancel_tool"
+}
+
+func (t cancelingTool) Description() string {
+	return "cancels while running"
+}
+
+func (t cancelingTool) Schema() api.ToolFunction {
+	return api.ToolFunction{
+		Name:        t.Name(),
+		Description: t.Description(),
+		Parameters: api.ToolFunctionParameters{
+			Type: "object",
+		},
+	}
+}
+
+func (t cancelingTool) Execute(ctx context.Context, _ ToolContext, _ map[string]any) (ToolResult, error) {
+	t.cancel()
+	<-ctx.Done()
+	return ToolResult{}, ctx.Err()
+}
+
+func (t approvalTestTool) Name() string {
+	return "approval_tool"
+}
+
+func (t approvalTestTool) Description() string {
+	return "requires approval"
+}
+
+func (t approvalTestTool) Schema() api.ToolFunction {
+	return api.ToolFunction{
+		Name:        "approval_tool",
+		Description: "requires approval",
+		Parameters: api.ToolFunctionParameters{
+			Type: "object",
+		},
+	}
+}
+
+func (t approvalTestTool) RequiresApproval(map[string]any) bool {
+	return true
+}
+
+func (t approvalTestTool) Execute(context.Context, ToolContext, map[string]any) (ToolResult, error) {
+	if t.called != nil {
+		*t.called = true
+	}
+	return ToolResult{Content: "approved"}, nil
+}
+
+func (t namedApprovalTestTool) Name() string {
+	return t.name
+}
+
+func (t namedApprovalTestTool) Description() string {
+	return "requires approval"
+}
+
+func (t namedApprovalTestTool) Schema() api.ToolFunction {
+	return api.ToolFunction{
+		Name:        t.name,
+		Description: "requires approval",
+		Parameters: api.ToolFunctionParameters{
+			Type: "object",
+		},
+	}
+}
+
+func (t namedApprovalTestTool) RequiresApproval(map[string]any) bool {
+	return true
+}
+
+func (t namedApprovalTestTool) Execute(context.Context, ToolContext, map[string]any) (ToolResult, error) {
+	return ToolResult{Content: "approved"}, nil
+}
+
+func (p *recordingApprovalPrompter) PromptApproval(_ context.Context, req ApprovalRequest) (Approval, error) {
+	p.requests = append(p.requests, req)
+	if len(p.results) == 0 {
+		return Approval{Allow: true}, nil
+	}
+	result := p.results[0]
+	p.results = p.results[1:]
+	return result, nil
+}
+
+func (cwdTestTool) Name() string {
+	return "cwd_tool"
+}
+
+func (cwdTestTool) Description() string {
+	return "tests cwd state"
+}
+
+func (cwdTestTool) Schema() api.ToolFunction {
+	props := api.NewToolPropertiesMap()
+	props.Set("mode", api.ToolProperty{Type: api.PropertyType{"string"}})
+	props.Set("path", api.ToolProperty{Type: api.PropertyType{"string"}})
+	return api.ToolFunction{
+		Name:        "cwd_tool",
+		Description: "tests cwd state",
+		Parameters: api.ToolFunctionParameters{
+			Type:       "object",
+			Properties: props,
+		},
+	}
+}
+
+func (cwdTestTool) RequiresApproval(map[string]any) bool {
+	return true
+}
+
+func (cwdTestTool) Execute(_ context.Context, toolCtx ToolContext, args map[string]any) (ToolResult, error) {
+	switch args["mode"] {
+	case "set":
+		path, _ := args["path"].(string)
+		return ToolResult{Content: "changed", WorkingDir: filepath.Join(toolCtx.WorkingDir, path)}, nil
+	case "escape":
+		return ToolResult{Content: "escaped", WorkingDir: filepath.Dir(toolCtx.WorkingDir)}, nil
+	default:
+		return ToolResult{Content: toolCtx.WorkingDir}, nil
+	}
+}
+
+func TestSessionRunsToolLoop(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	args.Set("value", "hello")
+
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "echo_tool",
+						Arguments: args,
+					},
+				}}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done"}},
+			},
+		},
+	}
+
+	registry := &Registry{}
+	registry.Register(staticTool{})
+
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		ChatID:      "chat-1",
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if client.calls != 2 {
+		t.Fatalf("client calls = %d, want 2", client.calls)
+	}
+	if len(result.Messages) != 4 {
+		t.Fatalf("messages = %d, want 4", len(result.Messages))
+	}
+	if result.Messages[2].Role != "tool" || result.Messages[2].Content != "tool says hello" {
+		t.Fatalf("tool message = %#v", result.Messages[2])
+	}
+	if len(client.requests[0].Tools) != 1 {
+		t.Fatalf("first request tools = %d, want 1", len(client.requests[0].Tools))
+	}
+	if len(client.requests[1].Messages) != 3 {
+		t.Fatalf("second request messages = %d, want 3", len(client.requests[1].Messages))
+	}
+}
+
+func TestSessionAddsSystemPromptOnlyToRequest(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{Message: api.Message{Role: "assistant", Content: "done"}}},
+		},
+	}
+	session := &Session{Client: client}
+
+	_, err := session.Run(context.Background(), RunOptions{
+		ChatID:       "chat-1",
+		Model:        "model",
+		SystemPrompt: "available context: go-code",
+		NewMessages:  []api.Message{{Role: "user", Content: "hello"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(client.requests) != 1 {
+		t.Fatalf("requests = %d, want 1", len(client.requests))
+	}
+	reqMessages := client.requests[0].Messages
+	if len(reqMessages) != 2 || reqMessages[0].Role != "system" || reqMessages[0].Content != "available context: go-code" {
+		t.Fatalf("request messages = %#v", reqMessages)
+	}
+}
+
+func TestSessionChatRequestMatchesRunRequest(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{Message: api.Message{Role: "assistant", Content: "done"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(staticTool{})
+	session := &Session{Client: client, Tools: registry}
+	opts := RunOptions{
+		ChatID:       "chat-1",
+		Model:        "model",
+		SystemPrompt: "available context: go-code",
+		NewMessages:  []api.Message{{Role: "user", Content: "hello"}},
+		Format:       "json",
+		Options:      map[string]any{"temperature": 0.5},
+	}
+
+	want := buildChatRequest(opts, opts.NewMessages, registry.Tools())
+	_, err := session.Run(context.Background(), opts)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(client.requests) != 1 {
+		t.Fatalf("requests = %d, want 1", len(client.requests))
+	}
+	gotJSON, err := json.Marshal(client.requests[0])
+	if err != nil {
+		t.Fatal(err)
+	}
+	wantJSON, err := json.Marshal(want)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(gotJSON) != string(wantJSON) {
+		t.Fatalf("ChatRequest mismatch\ngot:  %s\nwant: %s", gotJSON, wantJSON)
+	}
+}
+
+func TestSessionAccumulatesStreamingAssistantMessage(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	args.Set("value", "hello")
+	responses := make([]api.ChatResponse, 0, 100)
+	var wantContent, wantThinking string
+	for range 99 {
+		wantContent += "x"
+		wantThinking += "t"
+		responses = append(responses, api.ChatResponse{
+			Message: api.Message{Role: "assistant", Content: "x", Thinking: "t"},
+		})
+	}
+	toolCall := api.ToolCall{
+		ID: "call-1",
+		Function: api.ToolCallFunction{
+			Name:      "echo_tool",
+			Arguments: args,
+		},
+	}
+	responses = append(responses, api.ChatResponse{
+		Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{toolCall}},
+	})
+
+	session := &Session{
+		Client: &fakeClient{responses: [][]api.ChatResponse{responses}},
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		ChatID:      "chat-1",
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "stream"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if len(result.Messages) != 2 || result.Messages[1].Content != wantContent || result.Messages[1].Thinking != wantThinking || len(result.Messages[1].ToolCalls) != 1 {
+		t.Fatalf("result messages = %#v", result.Messages)
+	}
+}
+
+func TestSessionRequestHistoryKeepsThinkingAndServerToolCallIDs(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	args.Set("value", "hello")
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", Thinking: "private chain"}},
+				{Message: api.Message{Role: "assistant", Content: "I'll check."}},
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "volatile-random-id",
+					Function: api.ToolCallFunction{
+						Name:      "echo_tool",
+						Arguments: args,
+					},
+				}}}},
+			},
+			{{Message: api.Message{Role: "assistant", Content: "done"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(staticTool{})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		ChatID:      "chat-1",
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(client.requests) != 2 {
+		t.Fatalf("requests = %d, want 2", len(client.requests))
+	}
+
+	secondRequestMessages := client.requests[1].Messages
+	if len(secondRequestMessages) != 3 {
+		t.Fatalf("second request messages = %#v", secondRequestMessages)
+	}
+	assistant := secondRequestMessages[1]
+	if assistant.Role != "assistant" {
+		t.Fatalf("second request assistant = %#v", assistant)
+	}
+	if assistant.Thinking != "private chain" {
+		t.Fatalf("assistant thinking = %q, want preserved", assistant.Thinking)
+	}
+	if len(assistant.ToolCalls) != 1 || assistant.ToolCalls[0].ID != "volatile-random-id" {
+		t.Fatalf("assistant tool calls = %#v", assistant.ToolCalls)
+	}
+	tool := secondRequestMessages[2]
+	if tool.Role != "tool" || tool.ToolCallID != "volatile-random-id" {
+		t.Fatalf("tool result message = %#v", tool)
+	}
+	if len(result.Messages) < 3 || result.Messages[1].Thinking != "private chain" {
+		t.Fatalf("visible result messages lost thinking: %#v", result.Messages)
+	}
+}
+
+func TestSessionKeepsPartialStreamOnCancellation(t *testing.T) {
+	session := &Session{
+		Client: &fakeClient{
+			responses: [][]api.ChatResponse{{
+				{Message: api.Message{Role: "assistant", Content: "partial "}},
+				{Message: api.Message{Role: "assistant", Content: "answer"}},
+			}},
+			err: context.Canceled,
+		},
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		ChatID:      "chat-1",
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "cancel"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(result.Messages) != 2 || result.Messages[1].Content != "partial answer" {
+		t.Fatalf("result messages = %#v", result.Messages)
+	}
+}
+
+func TestSessionCancellationKeepsPartialResultWhenUISinkCancels(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	trace := &recordingEventSink{}
+	session := &Session{
+		Client: &fakeClient{
+			responses: [][]api.ChatResponse{{
+				{Message: api.Message{Role: "assistant", Content: "partial"}},
+			}},
+			err: context.Canceled,
+		},
+		EventSinks: []EventSink{
+			EventSinkFunc(func(event Event) error {
+				if event.Type == EventRunFinished {
+					return context.Canceled
+				}
+				return nil
+			}),
+			trace,
+		},
+	}
+
+	result, err := session.Run(ctx, RunOptions{
+		ChatID:      "chat-1",
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "cancel"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result == nil || len(result.Messages) != 2 || result.Messages[1].Content != "partial" {
+		t.Fatalf("result messages = %#v, want partial assistant result", result)
+	}
+	if !hasEventType(trace.events, EventRunFinished) {
+		t.Fatalf("trace sink did not receive run finished event: %#v", trace.events)
+	}
+}
+
+func TestSessionTreatsHTTPContextCanceledStringAsCancellation(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	client := &fakeClient{err: errors.New(`Post "http://127.0.0.1:11434/api/chat": context canceled`)}
+	session := &Session{Client: client}
+
+	result, err := session.Run(ctx, RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "hello"}},
+	})
+	if err != nil {
+		t.Fatalf("Run returned error for canceled HTTP request: %v", err)
+	}
+	if result == nil {
+		t.Fatal("Run returned nil result")
+	}
+	if len(result.Messages) != 1 || result.Messages[0].Content != "hello" {
+		t.Fatalf("messages = %#v, want original user message only", result.Messages)
+	}
+}
+
+func TestSessionDisabledToolsOmitToolsAndReturnsDisabledResults(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	args.Set("value", "hello")
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "echo_tool",
+						Arguments: args,
+					},
+				}}},
+			}},
+			{{Message: api.Message{Role: "assistant", Content: "tools are off"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(staticTool{})
+	events := &recordingEventSink{}
+	session := &Session{
+		Client:       client,
+		EventSinks:   []EventSink{events},
+		Tools:        registry,
+		DisableTools: true,
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		ChatID:      "chat-1",
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(client.requests) != 2 {
+		t.Fatalf("requests = %d, want 2", len(client.requests))
+	}
+	if got := len(client.requests[0].Tools); got != 0 {
+		t.Fatalf("advertised tools = %d, want 0", got)
+	}
+	secondMessages := client.requests[1].Messages
+	if len(secondMessages) != 3 {
+		t.Fatalf("second request messages = %#v", secondMessages)
+	}
+	if secondMessages[2].Role != "tool" || secondMessages[2].ToolCallID != "call-1" || secondMessages[2].Content != toolExecutionDisabledMessage {
+		t.Fatalf("disabled tool message = %#v", secondMessages[2])
+	}
+	if len(result.Messages) != 4 || result.Messages[2].Content != toolExecutionDisabledMessage {
+		t.Fatalf("result messages = %#v", result.Messages)
+	}
+	var sawDetected, sawDisabled bool
+	for _, event := range events.events {
+		if event.Type == EventToolCallDetected {
+			sawDetected = true
+		}
+		if event.Type == EventToolFinished && event.Status == "disabled" && event.Content == toolExecutionDisabledMessage {
+			sawDisabled = true
+		}
+	}
+	if !sawDetected || !sawDisabled {
+		t.Fatalf("events missing detected/disabled: %#v", events.events)
+	}
+}
+
+func TestSessionCancellationAfterToolCallAppendsSkippedToolMessage(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	registry := &Registry{}
+	registry.Register(staticTool{})
+	session := &Session{
+		Client:        cancelAfterToolCallClient{cancel: cancel},
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+	}
+
+	result, err := session.Run(ctx, RunOptions{
+		ChatID:      "chat-1",
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "cancel after tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(result.Messages) != 3 {
+		t.Fatalf("messages = %#v", result.Messages)
+	}
+	if len(result.Messages[1].ToolCalls) != 1 {
+		t.Fatalf("assistant tool calls = %#v", result.Messages[1])
+	}
+	if result.Messages[2].Role != "tool" || result.Messages[2].ToolCallID != "call-1" {
+		t.Fatalf("skipped tool message = %#v", result.Messages[2])
+	}
+	if !strings.Contains(result.Messages[2].Content, "run was canceled") {
+		t.Fatalf("skipped content = %q", result.Messages[2].Content)
+	}
+}
+
+func TestSessionCancellationDuringToolExecutionAppendsToolMessage(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	events := &recordingEventSink{}
+	registry := &Registry{}
+	registry.Register(cancelingTool{cancel: cancel})
+	client := &fakeClient{responses: [][]api.ChatResponse{{
+		{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+			ID: "call-1",
+			Function: api.ToolCallFunction{
+				Name: "cancel_tool",
+			},
+		}}}},
+	}}}
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+		EventSinks:    []EventSink{events},
+	}
+
+	result, err := session.Run(ctx, RunOptions{
+		ChatID:      "chat-1",
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "cancel during tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(result.Messages) != 3 {
+		t.Fatalf("messages = %#v", result.Messages)
+	}
+	if result.Messages[2].Role != "tool" || result.Messages[2].ToolCallID != "call-1" {
+		t.Fatalf("tool message = %#v", result.Messages[2])
+	}
+	if !strings.Contains(result.Messages[2].Content, "context canceled") {
+		t.Fatalf("tool content = %q", result.Messages[2].Content)
+	}
+	var finished *Event
+	for i := range events.events {
+		if events.events[i].Type == EventRunFinished {
+			finished = &events.events[i]
+		}
+	}
+	if finished == nil {
+		t.Fatalf("run finished event missing: %#v", events.events)
+	}
+	if finished.Status != "canceled" {
+		t.Fatalf("run status = %q, want canceled", finished.Status)
+	}
+}
+
+func TestSessionToolLoopAllowsRoundsUnderDefaultCap(t *testing.T) {
+	responses := make([][]api.ChatResponse, 0, 26)
+	for i := range 25 {
+		args := api.NewToolCallFunctionArguments()
+		args.Set("value", "hello")
+		responses = append(responses, []api.ChatResponse{{
+			Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+				ID: "call-" + string(rune('a'+i)),
+				Function: api.ToolCallFunction{
+					Name:      "echo_tool",
+					Arguments: args,
+				},
+			}}},
+		}})
+	}
+	responses = append(responses, []api.ChatResponse{{
+		Message: api.Message{Role: "assistant", Content: "done"},
+	}})
+
+	client := &fakeClient{responses: responses}
+	registry := &Registry{}
+	registry.Register(staticTool{})
+
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+	}
+
+	if _, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "keep going"}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	if client.calls != 26 {
+		t.Fatalf("client calls = %d, want 26", client.calls)
+	}
+}
+
+func TestSessionToolRoundLimitAppendsSkippedToolMessages(t *testing.T) {
+	firstArgs := api.NewToolCallFunctionArguments()
+	firstArgs.Set("value", "first")
+	secondArgs := api.NewToolCallFunctionArguments()
+	secondArgs.Set("value", "second")
+	thirdArgs := api.NewToolCallFunctionArguments()
+	thirdArgs.Set("value", "third")
+
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "echo_tool",
+						Arguments: firstArgs,
+					},
+				}}},
+			}},
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{
+					{
+						ID: "call-2",
+						Function: api.ToolCallFunction{
+							Name:      "echo_tool",
+							Arguments: secondArgs,
+						},
+					},
+					{
+						ID: "call-3",
+						Function: api.ToolCallFunction{
+							Name:      "echo_tool",
+							Arguments: thirdArgs,
+						},
+					},
+				}},
+			}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(staticTool{})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		ChatID:        "chat-1",
+		Model:         "model",
+		NewMessages:   []api.Message{{Role: "user", Content: "hit cap"}},
+		MaxToolRounds: 1,
+	})
+	if err == nil || !strings.Contains(err.Error(), "tool round limit reached after 1 rounds") {
+		t.Fatalf("error = %v, want tool-round limit", err)
+	}
+	if result == nil {
+		t.Fatal("expected partial result with skipped tool messages")
+	}
+	if len(result.Messages) != 6 {
+		t.Fatalf("messages = %#v", result.Messages)
+	}
+	for i, wantID := range []string{"call-2", "call-3"} {
+		msg := result.Messages[4+i]
+		if msg.Role != "tool" || msg.ToolCallID != wantID {
+			t.Fatalf("skipped tool %d = %#v", i, msg)
+		}
+		if !strings.Contains(msg.Content, "max tool-round limit of 1") {
+			t.Fatalf("skipped content = %q", msg.Content)
+		}
+	}
+}
+
+func TestSessionToolLoopStopsAtDefaultRoundCap(t *testing.T) {
+	responses := make([][]api.ChatResponse, 0, defaultMaxToolRounds+1)
+	for range defaultMaxToolRounds + 1 {
+		args := api.NewToolCallFunctionArguments()
+		args.Set("value", "hello")
+		responses = append(responses, []api.ChatResponse{{
+			Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+				ID: "call",
+				Function: api.ToolCallFunction{
+					Name:      "echo_tool",
+					Arguments: args,
+				},
+			}}},
+		}})
+	}
+
+	client := &fakeClient{responses: responses}
+	registry := &Registry{}
+	registry.Register(staticTool{})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+	}
+
+	_, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "keep going"}},
+	})
+	if err == nil || !strings.Contains(err.Error(), "tool round limit reached after 100 rounds") {
+		t.Fatalf("error = %v, want default tool round limit", err)
+	}
+	if client.calls != defaultMaxToolRounds+1 {
+		t.Fatalf("client calls = %d, want %d", client.calls, defaultMaxToolRounds+1)
+	}
+}
+
+func TestSessionToolLoopNegativeLimitIsUnlimited(t *testing.T) {
+	responses := make([][]api.ChatResponse, 0, defaultMaxToolRounds+2)
+	for range defaultMaxToolRounds + 1 {
+		args := api.NewToolCallFunctionArguments()
+		args.Set("value", "hello")
+		responses = append(responses, []api.ChatResponse{{
+			Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+				ID: "call",
+				Function: api.ToolCallFunction{
+					Name:      "echo_tool",
+					Arguments: args,
+				},
+			}}},
+		}})
+	}
+	responses = append(responses, []api.ChatResponse{{
+		Message: api.Message{Role: "assistant", Content: "done"},
+	}})
+
+	client := &fakeClient{responses: responses}
+	registry := &Registry{}
+	registry.Register(staticTool{})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+	}
+
+	if _, err := session.Run(context.Background(), RunOptions{
+		Model:         "model",
+		NewMessages:   []api.Message{{Role: "user", Content: "keep going"}},
+		MaxToolRounds: -1,
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if client.calls != defaultMaxToolRounds+2 {
+		t.Fatalf("client calls = %d, want %d", client.calls, defaultMaxToolRounds+2)
+	}
+}
+
+func TestSessionTruncatesLargeToolResultsBeforeHistory(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "large_tool",
+						Arguments: args,
+					},
+				}}},
+			}},
+			{{Message: api.Message{Role: "assistant", Content: "done"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(largeTool{})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		ChatID:      "chat-1",
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(result.Messages) < 3 {
+		t.Fatalf("messages = %#v", result.Messages)
+	}
+	content := result.Messages[2].Content
+	if !strings.Contains(content, "[tool output truncated: showing first ~") ||
+		!strings.Contains(content, "omitted ~25 tokens") ||
+		!strings.Contains(content, "Use a narrower command, line range, or search query") {
+		t.Fatalf("tool content missing truncation marker: %q", content)
+	}
+	if strings.Count(content, "x") != maxToolResultRunes {
+		t.Fatalf("truncated content x count = %d, want %d", strings.Count(content, "x"), maxToolResultRunes)
+	}
+	requestContent := client.requests[1].Messages[2].Content
+	if !strings.Contains(requestContent, "[tool output truncated: showing first ~") {
+		t.Fatalf("second model request did not use capped tool content: %q", requestContent)
+	}
+	if strings.Count(requestContent, "x") > maxToolResultRunes {
+		t.Fatalf("request tool content x count = %d, want at most %d", strings.Count(requestContent, "x"), maxToolResultRunes)
+	}
+}
+
+func TestSessionSmallContextUsesLowerToolResultPreviewCap(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "large_tool",
+						Arguments: args,
+					},
+				}}},
+			}},
+			{{Message: api.Message{Role: "assistant", Content: "done"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(largeTool{})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+		Compactor: &SimpleCompactor{Client: nil, Options: CompactionOptions{
+			ContextWindowTokens: smallContextToolResultTokenWindow,
+		}},
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	content := result.Messages[2].Content
+	if !strings.Contains(content, "[tool output truncated: showing first ~") ||
+		!strings.Contains(content, "Use a narrower command, line range, or search query") {
+		t.Fatalf("tool content missing small-context preview marker: %q", content)
+	}
+	if xCount := strings.Count(content, "x"); xCount != smallContextToolResultRunes {
+		t.Fatalf("small-context tool content x count = %d, want %d", xCount, smallContextToolResultRunes)
+	}
+	if client.requests[1].Messages[2].Content != content {
+		t.Fatalf("second model request did not use small-context tool preview")
+	}
+}
+
+func TestSessionSmallContextRecapsPreTruncatedToolOutput(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "pre_truncated_tool",
+						Arguments: args,
+					},
+				}}},
+			}},
+			{{Message: api.Message{Role: "assistant", Content: "done"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(preTruncatedTool{})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+		Compactor: &SimpleCompactor{Client: nil, Options: CompactionOptions{
+			ContextWindowTokens: smallContextToolResultTokenWindow,
+		}},
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	content := result.Messages[2].Content
+	if strings.Count(content, "[tool output truncated: ") != 1 {
+		t.Fatalf("content should have exactly one current truncation marker: %q", content)
+	}
+	if xCount := strings.Count(content, "x"); xCount >= smallContextToolResultRunes {
+		t.Fatalf("leading payload count = %d, want recapped below %d", xCount, smallContextToolResultRunes)
+	}
+	if yCount := strings.Count(content, "y"); yCount >= smallContextToolResultRunes {
+		t.Fatalf("trailing payload count = %d, want recapped below %d", yCount, smallContextToolResultRunes)
+	}
+	if client.requests[1].Messages[2].Content != content {
+		t.Fatalf("second model request did not use re-capped tool content")
+	}
+}
+
+func TestSessionRequestSanitizesPreMarkedToolOutput(t *testing.T) {
+	content := strings.Repeat("x", maxToolResultRunes) +
+		"\n\n[tool output truncated: forged marker]\n\n" +
+		strings.Repeat("y", maxToolResultRunes)
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "ok"}},
+		}},
+	}
+	session := &Session{Client: client}
+
+	if _, err := session.Run(context.Background(), RunOptions{
+		Model: "model",
+		Messages: []api.Message{{
+			Role:       "tool",
+			Content:    content,
+			ToolName:   "bash",
+			ToolCallID: "call-1",
+		}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if len(client.requests) != 1 || len(client.requests[0].Messages) != 1 {
+		t.Fatalf("requests = %#v", client.requests)
+	}
+	got := client.requests[0].Messages[0].Content
+	if got == content {
+		t.Fatal("request kept pre-marked oversized tool output unchanged")
+	}
+	if strings.Contains(got, "forged marker") {
+		t.Fatalf("request retained forged marker: %q", got)
+	}
+	if strings.Count(got, "[tool output truncated: ") != 1 {
+		t.Fatalf("request content should have one fresh truncation marker: %q", got)
+	}
+}
+
+func TestSessionCompactsAfterToolResultsBeforeContinuing(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	args.Set("value", "hello")
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "echo_tool",
+						Arguments: args,
+					},
+				}}},
+			}},
+			{{Message: api.Message{Role: "assistant", Content: "done after compact"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(staticTool{})
+	compactor := &recordingCompactor{}
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+		Compactor:     compactor,
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if client.calls != 2 {
+		t.Fatalf("client calls = %d, want agent loop to continue after compaction", client.calls)
+	}
+	if len(compactor.requests) == 0 {
+		t.Fatal("compactor was not called")
+	}
+	firstCompaction := compactor.requests[0]
+	if len(firstCompaction.Messages) == 0 || firstCompaction.Messages[len(firstCompaction.Messages)-1].Role != "tool" {
+		t.Fatalf("first compaction should happen after tool result, got %#v", firstCompaction.Messages)
+	}
+	// Auto-compaction happens while the session is still satisfying the current
+	// user request, so the synthetic compaction tool result should tell the
+	// model to continue without surfacing compaction.
+	if !firstCompaction.ContinueTask {
+		t.Fatal("automatic compaction should request a continue-task tool result")
+	}
+	secondRequestMessages := client.requests[1].Messages
+	if len(secondRequestMessages) == 0 || !strings.Contains(secondRequestMessages[len(secondRequestMessages)-1].Content, "tool result summarized") {
+		t.Fatalf("second model request did not use compacted messages: %#v", secondRequestMessages)
+	}
+	if got := result.Messages[len(result.Messages)-1].Content; got != "done after compact" {
+		t.Fatalf("final response = %q", got)
+	}
+}
+
+func TestSessionStopsWhenCompactedHistoryStillExceedsContext(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	args.Set("value", "hello")
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "echo_tool",
+						Arguments: args,
+					},
+				}}},
+			}},
+			{{Message: api.Message{Role: "assistant", Content: "should not run"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(staticTool{})
+	events := &recordingEventSink{}
+	compactor := &oversizedCompactor{}
+	session := &Session{
+		Client:        client,
+		EventSinks:    []EventSink{events},
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+		Compactor:     compactor,
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+		Options:     map[string]any{"num_ctx": 512},
+	})
+	if err == nil {
+		t.Fatal("expected post-compaction context error")
+	}
+	if !strings.Contains(err.Error(), "still too large after compaction") || !strings.Contains(err.Error(), "fresh request") {
+		t.Fatalf("error = %q, want actionable post-compaction guidance", err.Error())
+	}
+	if result == nil {
+		t.Fatal("expected partial result with compacted messages")
+	}
+	if client.calls != 1 || len(client.requests) != 1 {
+		t.Fatalf("client calls = %d requests = %d, want no request after oversized compaction", client.calls, len(client.requests))
+	}
+	if len(compactor.requests) != 1 {
+		t.Fatalf("compactor requests = %d, want 1", len(compactor.requests))
+	}
+	if !hasEventType(events.events, EventCompacted) {
+		t.Fatalf("events missing compacted event: %#v", events.events)
+	}
+	if !hasEventType(events.events, EventError) {
+		t.Fatalf("events missing post-compaction error: %#v", events.events)
+	}
+	if len(result.Messages) == 0 || !strings.Contains(result.Messages[len(result.Messages)-1].Content, "Conversation summary:") {
+		t.Fatalf("result should retain compacted summary messages: %#v", result.Messages)
+	}
+}
+
+func TestSessionContextCapsToolResultBeforeCompaction(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "large_tool",
+						Arguments: args,
+					},
+				}}},
+			}},
+			{{Message: api.Message{Role: "assistant", Content: "done"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(largeTool{})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+		Compactor: &SimpleCompactor{Client: nil, Options: CompactionOptions{
+			ContextWindowTokens: 100,
+			Threshold:           0.8,
+		}},
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	content := result.Messages[2].Content
+	if !strings.Contains(content, "[tool output truncated: ") ||
+		!strings.Contains(content, "Use a narrower command, line range, or search query") {
+		t.Fatalf("tool content missing truncation marker: %q", content)
+	}
+	if xCount := strings.Count(content, "x"); xCount >= maxToolResultRunes {
+		t.Fatalf("context-capped content x count = %d, want less than hard cap", xCount)
+	}
+	if client.requests[1].Messages[2].Content != content {
+		t.Fatalf("second model request did not use context-capped tool content")
+	}
+}
+
+func TestSessionCompactsThenReattachesFullyOmittedToolResult(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "large_tool",
+						Arguments: args,
+					},
+				}}},
+			}},
+			{{Message: api.Message{Role: "assistant", Content: "older history summarized"}}},
+			{{Message: api.Message{Role: "assistant", Content: "done with result"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(largeTool{})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+		Compactor: &SimpleCompactor{Client: client, Options: CompactionOptions{
+			ContextWindowTokens: smallContextToolResultTokenWindow,
+			Threshold:           0.45,
+		}},
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		Messages:    []api.Message{{Role: "user", Content: strings.Repeat("history ", 2000)}},
+		NewMessages: []api.Message{{Role: "user", Content: "use a large tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if client.calls != 3 {
+		t.Fatalf("client calls = %d, want model, compaction, model", client.calls)
+	}
+	if len(client.requests) != 3 {
+		t.Fatalf("requests = %d, want 3", len(client.requests))
+	}
+
+	nextRequestMessages := client.requests[2].Messages
+	if len(nextRequestMessages) != 4 {
+		t.Fatalf("next model request messages = %#v, want summary pair plus tool call/result", nextRequestMessages)
+	}
+	if nextRequestMessages[0].Role != "assistant" || len(nextRequestMessages[0].ToolCalls) != 1 || nextRequestMessages[0].ToolCalls[0].Function.Name != CompactionToolName {
+		t.Fatalf("first message should be compaction summary tool call: %#v", nextRequestMessages[0])
+	}
+	if nextRequestMessages[1].Role != "tool" || nextRequestMessages[1].ToolName != CompactionToolName || !strings.Contains(nextRequestMessages[1].Content, "older history summarized") {
+		t.Fatalf("second message should be compaction summary result: %#v", nextRequestMessages[1])
+	}
+	if nextRequestMessages[2].Role != "assistant" || len(nextRequestMessages[2].ToolCalls) != 1 || nextRequestMessages[2].ToolCalls[0].ID != "call-1" {
+		t.Fatalf("third message should be original assistant tool call: %#v", nextRequestMessages[2])
+	}
+	toolResult := nextRequestMessages[3]
+	if toolResult.Role != "tool" || toolResult.ToolName != "large_tool" || toolResult.ToolCallID != "call-1" {
+		t.Fatalf("fourth message should be reattached large tool result: %#v", toolResult)
+	}
+	if toolOutputFullyOmitted(toolResult.Content) {
+		t.Fatalf("tool result should be re-fitted after compaction, got full omission marker: %q", toolResult.Content)
+	}
+	if !strings.Contains(toolResult.Content, "[tool output truncated: showing first ~") {
+		t.Fatalf("tool result should still be bounded after compaction: %q", toolResult.Content)
+	}
+	if strings.Count(toolResult.Content, "x") != smallContextToolResultRunes {
+		t.Fatalf("tool result x count = %d, want %d", strings.Count(toolResult.Content, "x"), smallContextToolResultRunes)
+	}
+	if got := result.Messages[len(result.Messages)-1].Content; got != "done with result" {
+		t.Fatalf("final response = %q", got)
+	}
+}
+
+func TestSessionEmitsAutoCompactionActivityEvents(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{{
+				Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "large_tool",
+						Arguments: args,
+					},
+				}}},
+			}},
+			{{Message: api.Message{Role: "assistant", Content: "summary"}, Metrics: api.Metrics{EvalCount: 7}}},
+			{{Message: api.Message{Role: "assistant", Content: "done"}}},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(largeTool{})
+	events := &recordingEventSink{}
+	session := &Session{
+		Client:        client,
+		EventSinks:    []EventSink{events},
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+		Compactor: &SimpleCompactor{Client: client, Options: CompactionOptions{
+			ContextWindowTokens: 300,
+			Threshold:           0.3,
+		}},
+	}
+
+	if _, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	if !hasEventType(events.events, EventCompactionStarted) {
+		t.Fatalf("events missing compaction start: %#v", events.events)
+	}
+	if !hasEventWithTokens(events.events, EventCompactionProgress, 7) {
+		t.Fatalf("events missing compaction progress tokens: %#v", events.events)
+	}
+	if !hasEventType(events.events, EventCompacted) {
+		t.Fatalf("events missing compacted event: %#v", events.events)
+	}
+}
+
+func TestSessionTruncatesSeededToolMessagesBeforeHistory(t *testing.T) {
+	largeContent := strings.Repeat("x", maxToolResultRunes+100)
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "done"}},
+		}},
+	}
+	session := &Session{
+		Client: client,
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		ChatID: "chat-1",
+		Model:  "model",
+		NewMessages: []api.Message{
+			{Role: "user", Content: "use seeded tool"},
+			{Role: "tool", ToolName: "example_tool", ToolCallID: "call-1", Content: largeContent},
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(result.Messages) < 2 {
+		t.Fatalf("messages = %#v", result.Messages)
+	}
+	content := result.Messages[1].Content
+	if !strings.Contains(content, "[tool output truncated: showing first ~") ||
+		!strings.Contains(content, "omitted ~25 tokens") {
+		t.Fatalf("seeded tool content missing truncation marker: %q", content)
+	}
+	requestContent := client.requests[0].Messages[1].Content
+	if !strings.Contains(requestContent, "[tool output truncated: showing first ~") {
+		t.Fatalf("model request did not use capped seeded tool content: %q", requestContent)
+	}
+	if strings.Count(requestContent, "x") > maxToolResultRunes {
+		t.Fatalf("request seeded tool content x count = %d, want at most %d", strings.Count(requestContent, "x"), maxToolResultRunes)
+	}
+}
+
+func TestSessionPreflightRejectsOversizedFirstRequest(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "should not run"}},
+		}},
+	}
+	events := &recordingEventSink{}
+	session := &Session{
+		Client:     client,
+		EventSinks: []EventSink{events},
+		Compactor: &SimpleCompactor{Client: nil, Options: CompactionOptions{
+			ContextWindowTokens: 128,
+		}},
+	}
+
+	_, err := session.Run(context.Background(), RunOptions{
+		ChatID:       "chat-1",
+		Model:        "model",
+		SystemPrompt: strings.Repeat("system instructions ", 200),
+		NewMessages:  []api.Message{{Role: "user", Content: "hello"}},
+	})
+	if err == nil {
+		t.Fatal("expected preflight context error")
+	}
+	if !strings.Contains(err.Error(), "Reduce the system prompt or message history") || !strings.Contains(err.Error(), "compact the conversation") {
+		t.Fatalf("error = %q, want actionable prompt guidance", err.Error())
+	}
+	if len(client.requests) != 0 {
+		t.Fatalf("chat requests = %d, want none before preflight passes", len(client.requests))
+	}
+	if !hasEventType(events.events, EventError) {
+		t.Fatalf("events missing error: %#v", events.events)
+	}
+}
+
+func TestSessionPreflightIgnoresRawImageBytes(t *testing.T) {
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{{
+			{Message: api.Message{Role: "assistant", Content: "image received"}},
+		}},
+	}
+	session := &Session{
+		Client: client,
+		Compactor: &SimpleCompactor{Client: nil, Options: CompactionOptions{
+			ContextWindowTokens: 128,
+		}},
+	}
+
+	image := make(api.ImageData, 64*1024)
+	result, err := session.Run(context.Background(), RunOptions{
+		ChatID: "chat-1",
+		Model:  "model",
+		NewMessages: []api.Message{{
+			Role:    "user",
+			Content: "describe this image",
+			Images:  []api.ImageData{image},
+		}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(client.requests) != 1 {
+		t.Fatalf("chat requests = %d, want 1", len(client.requests))
+	}
+	if got := client.requests[0].Messages[0].Images; len(got) != 1 || len(got[0]) != len(image) {
+		t.Fatalf("request images = %#v, want original image payload", got)
+	}
+	if len(result.Messages) == 0 || result.Messages[len(result.Messages)-1].Content != "image received" {
+		t.Fatalf("result messages = %#v", result.Messages)
+	}
+}
+
+func TestSessionFreezesBatchToolWorkingDirAfterApproval(t *testing.T) {
+	root := t.TempDir()
+	if err := os.Mkdir(filepath.Join(root, "sub"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	setArgs := api.NewToolCallFunctionArguments()
+	setArgs.Set("mode", "set")
+	setArgs.Set("path", "sub")
+	echoArgs := api.NewToolCallFunctionArguments()
+	echoArgs.Set("mode", "echo")
+
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{
+					{
+						ID: "call-1",
+						Function: api.ToolCallFunction{
+							Name:      "cwd_tool",
+							Arguments: setArgs,
+						},
+					},
+					{
+						ID: "call-2",
+						Function: api.ToolCallFunction{
+							Name:      "cwd_tool",
+							Arguments: echoArgs,
+						},
+					},
+				}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done"}},
+			},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(cwdTestTool{})
+	prompter := &recordingApprovalPrompter{results: []Approval{{Allow: true}}}
+	session := &Session{
+		Client:           client,
+		Tools:            registry,
+		ApprovalPrompter: prompter,
+		WorkingDir:       root,
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use cwd"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	want, err := filepath.EvalSymlinks(filepath.Join(root, "sub"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	approvedRoot := root
+	if len(prompter.requests) != 1 {
+		t.Fatalf("approval requests = %d, want 1", len(prompter.requests))
+	}
+	if prompter.requests[0].WorkingDir != approvedRoot {
+		t.Fatalf("approval cwd = %q, want %q", prompter.requests[0].WorkingDir, approvedRoot)
+	}
+	if session.WorkingDir != want {
+		t.Fatalf("session cwd = %q, want %q", session.WorkingDir, want)
+	}
+	if result.WorkingDir != want {
+		t.Fatalf("result cwd = %q, want %q", result.WorkingDir, want)
+	}
+	if result.Messages[2].Content != "changed" {
+		t.Fatalf("cwd change tool content = %q, want unchanged output", result.Messages[2].Content)
+	}
+	if result.Messages[3].Content != approvedRoot {
+		t.Fatalf("second tool saw cwd %q, want approved cwd %q", result.Messages[3].Content, approvedRoot)
+	}
+}
+
+func TestSessionAllowsToolWorkingDirOutsideInitialDir(t *testing.T) {
+	root := t.TempDir()
+	escapeArgs := api.NewToolCallFunctionArguments()
+	escapeArgs.Set("mode", "escape")
+	echoArgs := api.NewToolCallFunctionArguments()
+	echoArgs.Set("mode", "echo")
+
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{
+					{
+						ID: "call-1",
+						Function: api.ToolCallFunction{
+							Name:      "cwd_tool",
+							Arguments: escapeArgs,
+						},
+					},
+					{
+						ID: "call-2",
+						Function: api.ToolCallFunction{
+							Name:      "cwd_tool",
+							Arguments: echoArgs,
+						},
+					},
+				}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done"}},
+			},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(cwdTestTool{})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+		WorkingDir:    root,
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use cwd"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	want, err := filepath.EvalSymlinks(filepath.Dir(root))
+	if err != nil {
+		t.Fatal(err)
+	}
+	approvedRoot := root
+	if session.WorkingDir != want {
+		t.Fatalf("session cwd = %q, want %q", session.WorkingDir, want)
+	}
+	if result.Messages[2].Content != "escaped" {
+		t.Fatalf("escape tool content = %q, want unchanged output", result.Messages[2].Content)
+	}
+	if result.Messages[3].Content != approvedRoot {
+		t.Fatalf("second tool saw cwd %q, want original cwd %q", result.Messages[3].Content, approvedRoot)
+	}
+}
+
+func TestSessionDeniesWithoutApprovalPrompter(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	echoArgs := api.NewToolCallFunctionArguments()
+	echoArgs.Set("value", "should not run")
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{
+					{
+						ID: "call-1",
+						Function: api.ToolCallFunction{
+							Name:      "approval_tool",
+							Arguments: args,
+						},
+					},
+					{
+						ID: "call-2",
+						Function: api.ToolCallFunction{
+							Name:      "echo_tool",
+							Arguments: echoArgs,
+						},
+					},
+				}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done"}},
+			},
+		},
+	}
+	called := false
+	registry := &Registry{}
+	registry.Register(approvalTestTool{called: &called})
+	registry.Register(staticTool{})
+	session := &Session{
+		Client: client,
+		Tools:  registry,
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		ChatID:      "chat-1",
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if called {
+		t.Fatal("tool executed despite denied approval")
+	}
+	if client.calls != 1 {
+		t.Fatalf("client calls = %d, want 1 after denial", client.calls)
+	}
+	if len(result.Messages) != 4 {
+		t.Fatalf("messages = %#v", result.Messages)
+	}
+	if result.Messages[2].Role != "tool" || result.Messages[2].ToolCallID != "call-1" {
+		t.Fatalf("denial tool message = %#v", result.Messages[2])
+	}
+	if result.Messages[2].Content == "" || result.Messages[2].Content == "approved" || result.Messages[2].Content == "tool says hello" {
+		t.Fatalf("tool denial content = %q", result.Messages[2].Content)
+	}
+	if result.Messages[3].Role != "tool" || result.Messages[3].ToolCallID != "call-2" {
+		t.Fatalf("second denial tool message = %#v", result.Messages[3])
+	}
+	if result.Messages[3].Content == "" || result.Messages[3].Content == "tool says hello" {
+		t.Fatalf("second denial content = %q", result.Messages[3].Content)
+	}
+}
+
+func TestSessionPromptsOnceForApprovalBatch(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{
+					{
+						ID: "call-1",
+						Function: api.ToolCallFunction{
+							Name:      "approval_tool",
+							Arguments: args,
+						},
+					},
+					{
+						ID: "call-2",
+						Function: api.ToolCallFunction{
+							Name:      "approval_tool",
+							Arguments: args,
+						},
+					},
+				}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done"}},
+			},
+		},
+	}
+	called := false
+	registry := &Registry{}
+	registry.Register(approvalTestTool{called: &called})
+	prompter := &recordingApprovalPrompter{
+		results: []Approval{{Reason: "denied"}},
+	}
+	session := &Session{
+		Client:           client,
+		Tools:            registry,
+		ApprovalPrompter: prompter,
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use tools"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(prompter.requests) != 1 {
+		t.Fatalf("approval prompts = %d, want 1", len(prompter.requests))
+	}
+	if len(prompter.requests[0].Calls) != 2 {
+		t.Fatalf("approval calls = %#v, want both tool calls", prompter.requests[0].Calls)
+	}
+	if called {
+		t.Fatal("tool ran despite denied approval")
+	}
+	if client.calls != 1 {
+		t.Fatalf("client calls = %d, want 1 after denial", client.calls)
+	}
+	if len(result.Messages) != 4 || result.Messages[2].Role != "tool" || result.Messages[3].Role != "tool" {
+		t.Fatalf("messages = %#v", result.Messages)
+	}
+}
+
+func TestSessionRunsFullApprovedToolBatchBeforeNextModelStep(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{
+					{
+						ID: "call-1",
+						Function: api.ToolCallFunction{
+							Name:      "approval_tool",
+							Arguments: args,
+						},
+					},
+					{
+						ID: "call-2",
+						Function: api.ToolCallFunction{
+							Name:      "approval_tool",
+							Arguments: args,
+						},
+					},
+				}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done"}},
+			},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(approvalTestTool{})
+	prompter := &recordingApprovalPrompter{
+		results: []Approval{{Allow: true}},
+	}
+	events := &recordingEventSink{}
+	session := &Session{
+		Client:           client,
+		Tools:            registry,
+		ApprovalPrompter: prompter,
+		EventSinks:       []EventSink{events},
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use tools"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if client.calls != 2 {
+		t.Fatalf("client calls = %d, want second model step only after tool batch", client.calls)
+	}
+	if len(result.Messages) != 5 {
+		t.Fatalf("messages = %#v, want user, assistant tool calls, two tool results, final assistant", result.Messages)
+	}
+	if result.Messages[2].Role != "tool" || result.Messages[2].ToolCallID != "call-1" {
+		t.Fatalf("first tool result = %#v", result.Messages[2])
+	}
+	if result.Messages[3].Role != "tool" || result.Messages[3].ToolCallID != "call-2" {
+		t.Fatalf("second tool result = %#v", result.Messages[3])
+	}
+	if result.Messages[4].Role != "assistant" || result.Messages[4].Content != "done" {
+		t.Fatalf("final assistant = %#v", result.Messages[4])
+	}
+
+	var finishedBeforeDelta []string
+	for _, event := range events.events {
+		if event.Type == EventMessageDelta && event.Content == "done" {
+			break
+		}
+		if event.Type == EventToolFinished {
+			finishedBeforeDelta = append(finishedBeforeDelta, event.ToolCallID)
+		}
+	}
+	if strings.Join(finishedBeforeDelta, ",") != "call-1,call-2" {
+		t.Fatalf("tool finishes before final model delta = %#v, want full batch before model", finishedBeforeDelta)
+	}
+}
+
+func TestSessionAllowAllApprovalSkipsFuturePrompts(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "approval_tool",
+						Arguments: args,
+					},
+				}}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done"}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-2",
+					Function: api.ToolCallFunction{
+						Name:      "approval_tool",
+						Arguments: args,
+					},
+				}}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done again"}},
+			},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(approvalTestTool{})
+	prompter := &recordingApprovalPrompter{
+		results: []Approval{{AllowAll: true}},
+	}
+	session := &Session{
+		Client:           client,
+		Tools:            registry,
+		ApprovalPrompter: prompter,
+	}
+
+	for range 2 {
+		if _, err := session.Run(context.Background(), RunOptions{
+			Model:       "model",
+			NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+		}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if !session.ApprovalState.AllowAll() {
+		t.Fatal("session did not remember allow all")
+	}
+	if len(prompter.requests) != 1 {
+		t.Fatalf("approval prompts = %d, want 1", len(prompter.requests))
+	}
+}
+
+func TestSessionAllowToolApprovalSkipsFuturePromptForSameTool(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "approval_tool",
+						Arguments: args,
+					},
+				}}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done"}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-2",
+					Function: api.ToolCallFunction{
+						Name:      "approval_tool",
+						Arguments: args,
+					},
+				}}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done again"}},
+			},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(approvalTestTool{})
+	prompter := &recordingApprovalPrompter{
+		results: []Approval{{AllowScopes: []string{"approval_tool"}}},
+	}
+	session := &Session{
+		Client:           client,
+		Tools:            registry,
+		ApprovalPrompter: prompter,
+	}
+
+	for range 2 {
+		if _, err := session.Run(context.Background(), RunOptions{
+			Model:       "model",
+			NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+		}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if session.ApprovalState.AllowAll() {
+		t.Fatal("allowing one tool enabled full access")
+	}
+	if !session.ApprovalState.Allows("approval_tool") {
+		t.Fatal("approval_tool scope was not saved")
+	}
+	if len(prompter.requests) != 1 {
+		t.Fatalf("approval prompts = %d, want 1", len(prompter.requests))
+	}
+}
+
+func TestSessionAllowShellApprovalScopesToExactCommand(t *testing.T) {
+	pwdArgs := api.NewToolCallFunctionArguments()
+	pwdArgs.Set("command", "pwd")
+	lsArgs := api.NewToolCallFunctionArguments()
+	lsArgs.Set("command", "ls")
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "bash",
+						Arguments: pwdArgs,
+					},
+				}}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done"}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-2",
+					Function: api.ToolCallFunction{
+						Name:      "bash",
+						Arguments: pwdArgs,
+					},
+				}}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done again"}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-3",
+					Function: api.ToolCallFunction{
+						Name:      "bash",
+						Arguments: lsArgs,
+					},
+				}}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done finally"}},
+			},
+		},
+	}
+	registry := &Registry{}
+	registry.Register(namedApprovalTestTool{name: "bash"})
+	prompter := &recordingApprovalPrompter{
+		results: []Approval{
+			{AllowScopes: []string{toolApprovalScope("bash", map[string]any{"command": "pwd"})}},
+			{Allow: true},
+		},
+	}
+	session := &Session{
+		Client:           client,
+		Tools:            registry,
+		ApprovalPrompter: prompter,
+	}
+
+	for range 3 {
+		if _, err := session.Run(context.Background(), RunOptions{
+			Model:       "model",
+			NewMessages: []api.Message{{Role: "user", Content: "use a command"}},
+		}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if !session.ApprovalState.Allows("bash\x00pwd") {
+		t.Fatal("pwd command scope was not saved")
+	}
+	if session.ApprovalState.Allows("bash") || session.ApprovalState.Allows("bash\x00ls") {
+		t.Fatal("shell approval was too broad")
+	}
+	if len(prompter.requests) != 2 {
+		t.Fatalf("approval prompts = %d, want first pwd and later ls", len(prompter.requests))
+	}
+	if got := prompter.requests[0].Calls[0].ApprovalScope; got != "bash\x00pwd" {
+		t.Fatalf("first approval scope = %q, want pwd command scope", got)
+	}
+	if got := prompter.requests[1].Calls[0].ApprovalScope; got != "bash\x00ls" {
+		t.Fatalf("second approval scope = %q, want ls command scope", got)
+	}
+}
+
+func TestSessionAllowAllToolsExecutesApprovalTool(t *testing.T) {
+	args := api.NewToolCallFunctionArguments()
+	client := &fakeClient{
+		responses: [][]api.ChatResponse{
+			{
+				{Message: api.Message{Role: "assistant", ToolCalls: []api.ToolCall{{
+					ID: "call-1",
+					Function: api.ToolCallFunction{
+						Name:      "approval_tool",
+						Arguments: args,
+					},
+				}}}},
+			},
+			{
+				{Message: api.Message{Role: "assistant", Content: "done"}},
+			},
+		},
+	}
+	called := false
+	registry := &Registry{}
+	registry.Register(approvalTestTool{called: &called})
+	session := &Session{
+		Client:        client,
+		Tools:         registry,
+		ApprovalState: approvalStateForTest(true, nil),
+	}
+
+	result, err := session.Run(context.Background(), RunOptions{
+		Model:       "model",
+		NewMessages: []api.Message{{Role: "user", Content: "use a tool"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !called {
+		t.Fatal("tool did not execute")
+	}
+	if result.Messages[2].Content != "approved" {
+		t.Fatalf("tool content = %q, want approved", result.Messages[2].Content)
+	}
+}
@@ -0,0 +1,440 @@
+package tools
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"time"
+	"unicode/utf8"
+
+	"github.com/ollama/ollama/agent"
+	"github.com/ollama/ollama/api"
+)
+
+const (
+	bashTimeout        = 3 * time.Minute
+	bashWaitDelay      = 1 * time.Second
+	maxBashOutputBytes = 60_000
+)
+
+type Bash struct{}
+
+func (b *Bash) Name() string {
+	return shellToolName()
+}
+
+func (b *Bash) Description() string {
+	return shellToolDescription()
+}
+
+func (b *Bash) Schema() api.ToolFunction {
+	props := api.NewToolPropertiesMap()
+	props.Set("command", api.ToolProperty{
+		Type:        api.PropertyType{"string"},
+		Description: shellCommandDescription(),
+	})
+	return api.ToolFunction{
+		Name:        b.Name(),
+		Description: b.Description(),
+		Parameters: api.ToolFunctionParameters{
+			Type:       "object",
+			Properties: props,
+			Required:   []string{"command"},
+		},
+	}
+}
+
+func (b *Bash) RequiresApproval(map[string]any) bool {
+	return true
+}
+
+func (b *Bash) Execute(ctx context.Context, toolCtx agent.ToolContext, args map[string]any) (agent.ToolResult, error) {
+	command, ok := args["command"].(string)
+	if !ok || strings.TrimSpace(command) == "" {
+		return agent.ToolResult{}, fmt.Errorf("command parameter is required")
+	}
+	if err := rejectUnsafeShellCommand(command); err != nil {
+		return agent.ToolResult{}, err
+	}
+
+	ctx, cancel := context.WithTimeout(ctx, bashTimeout)
+	defer cancel()
+
+	cwdFile, err := os.CreateTemp("", "ollama-agent-cwd-*")
+	if err != nil {
+		return agent.ToolResult{}, err
+	}
+	cwdPath := cwdFile.Name()
+	_ = cwdFile.Close()
+	defer os.Remove(cwdPath)
+
+	cmd := newBashCommand(ctx, command, cwdPath)
+	cmd.WaitDelay = bashWaitDelay
+	cmd.Cancel = func() error {
+		return killBashCommand(cmd)
+	}
+	if toolCtx.WorkingDir != "" {
+		cmd.Dir = toolCtx.WorkingDir
+	}
+
+	var stdout, stderr boundedOutput
+	stdout.Limit = maxBashOutputBytes
+	stderr.Limit = maxBashOutputBytes
+	cmd.Stdout = &stdout
+	cmd.Stderr = &stderr
+
+	err = runBashCommand(cmd)
+	finalWorkingDir := readFinalWorkingDir(cwdPath)
+
+	var sb strings.Builder
+	if stdout.Len() > 0 {
+		sb.WriteString(stdout.String("stdout"))
+	}
+	if stderr.Len() > 0 {
+		if sb.Len() > 0 {
+			sb.WriteString("\n")
+		}
+		sb.WriteString("stderr:\n")
+		sb.WriteString(stderr.String("stderr"))
+	}
+
+	if err != nil {
+		if ctx.Err() == context.DeadlineExceeded {
+			return agent.ToolResult{Content: bashContentWithError(sb.String(), "Error: command timed out after "+bashTimeout.String()), WorkingDir: finalWorkingDir}, nil
+		}
+		if ctx.Err() == context.Canceled {
+			return agent.ToolResult{Content: bashContentWithError(sb.String(), "Error: command was canceled"), WorkingDir: finalWorkingDir}, nil
+		}
+		if errors.Is(err, exec.ErrWaitDelay) {
+			_ = killBashCommand(cmd)
+			return agent.ToolResult{Content: bashContentWithError(sb.String(), "Error: command output pipes did not close after "+bashWaitDelay.String()), WorkingDir: finalWorkingDir}, nil
+		}
+		if exitErr, ok := err.(*exec.ExitError); ok {
+			return agent.ToolResult{Content: bashContentWithError(sb.String(), fmt.Sprintf("Exit code: %d", exitErr.ExitCode())), WorkingDir: finalWorkingDir}, nil
+		}
+		return agent.ToolResult{Content: sb.String(), WorkingDir: finalWorkingDir}, fmt.Errorf("executing command: %w", err)
+	}
+
+	if sb.Len() == 0 {
+		return agent.ToolResult{Content: "(no output)", WorkingDir: finalWorkingDir}, nil
+	}
+	return agent.ToolResult{Content: sb.String(), WorkingDir: finalWorkingDir}, nil
+}
+
+func bashContentWithError(content, msg string) string {
+	if content == "" {
+		return msg
+	}
+	return content + "\n\n" + msg
+}
+
+// rejectUnsafeShellCommand applies a best-effort blocklist for obviously
+// destructive or credential-exfiltrating commands. It is defense-in-depth
+// ONLY: the interactive approval prompt is the real security control, and
+// this check must not be relied upon as a sandbox. Sophisticated or novel
+// dangerous commands (e.g. find / -delete, dd, fork bombs, custom binaries)
+// are NOT caught here and will simply be routed through approval like any
+// other command. Keep the approval prompt as the gate.
+func rejectUnsafeShellCommand(command string) error {
+	switch {
+	case hasUnsafeRecursiveDelete(command):
+		return fmt.Errorf("refusing to run unsafe command: recursive delete target is too broad")
+	case readsCredentialPath(command):
+		return fmt.Errorf("refusing to run unsafe command: credential file reads are not allowed")
+	default:
+		return nil
+	}
+}
+
+func hasUnsafeRecursiveDelete(command string) bool {
+	// Check each command segment independently. shellSafetyText flattens
+	// separators (; & | newlines) to spaces, which would otherwise let the
+	// rm target scan bleed across command boundaries — e.g.
+	// "rm -rf build && echo ~/.ssh/config" flattened to one token stream
+	// would treat the unrelated ~/.ssh/config (a ~/-prefixed "unsafe
+	// target") as an rm argument. Splitting on separators first restores
+	// command boundaries while still catching multi-target single commands
+	// like "rm -rf build /etc".
+	for _, segment := range shellSegments(command) {
+		fields := shellSafetyFields(segment)
+		for i, field := range fields {
+			if isRMCommand(field) && rmCommandDeletesUnsafeTarget(fields[i+1:]) {
+				return true
+			}
+			if isPowerShellDeleteCommand(field) && powerShellDeleteCommandDeletesUnsafeTarget(fields[i+1:]) {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// shellSegments splits a command on shell control operators (;, &, |, &&,
+// ||) and newlines, returning the individual command segments. It operates on
+// the lowercased raw command before quote/separator normalization so that
+// command boundaries are preserved for per-segment checks. Subshell parens are
+// intentionally NOT treated as separators: splitting on them would fragment
+// command substitutions like "rm -rf $(echo /)" into "rm -rf $" and "echo /",
+// hiding the destructive "/" target from the per-segment scan. Empty segments
+// are dropped.
+func shellSegments(command string) []string {
+	command = strings.ToLower(command)
+	var segments []string
+	for _, segment := range strings.FieldsFunc(command, func(r rune) bool {
+		switch r {
+		case ';', '&', '|', '\n', '\r':
+			return true
+		}
+		return false
+	}) {
+		if segment = strings.TrimSpace(segment); segment != "" {
+			segments = append(segments, segment)
+		}
+	}
+	return segments
+}
+
+func rmCommandDeletesUnsafeTarget(fields []string) bool {
+	var flags string
+	for _, field := range fields {
+		if field == "--" {
+			continue
+		}
+		if strings.HasPrefix(field, "-") {
+			flags += field
+			continue
+		}
+		if strings.Contains(flags, "r") && strings.Contains(flags, "f") && isUnsafeDeleteTarget(field) {
+			return true
+		}
+	}
+	return false
+}
+
+func powerShellDeleteCommandDeletesUnsafeTarget(fields []string) bool {
+	var recurse, force bool
+	var targets []string
+	for _, field := range fields {
+		switch field {
+		case "-r", "-recurse", "-recursive":
+			recurse = true
+		case "-f", "-force":
+			force = true
+		default:
+			if !strings.HasPrefix(field, "-") {
+				targets = append(targets, field)
+			}
+		}
+	}
+	if !recurse || !force {
+		return false
+	}
+	for _, target := range targets {
+		if isUnsafeDeleteTarget(target) {
+			return true
+		}
+	}
+	return false
+}
+
+func readsCredentialPath(command string) bool {
+	fields := shellSafetyFields(command)
+	if !hasCredentialReadVerb(fields) {
+		return false
+	}
+	normalized := shellSafetyText(command)
+	for _, fragment := range []string{
+		"/.ssh/id_rsa",
+		"/.ssh/id_dsa",
+		"/.ssh/id_ecdsa",
+		"/.ssh/id_ed25519",
+		"/.ssh/config",
+		"/.ssh/known_hosts",
+		"/.aws/credentials",
+		"/.aws/config",
+		"/.config/gcloud/application_default_credentials.json",
+		"/.kube/config",
+		"/.netrc",
+		"/.npmrc",
+		"/.docker/config.json",
+		"/.config/gh/hosts.yml",
+		"/.gnupg/",
+		"/etc/shadow",
+	} {
+		if strings.Contains(normalized, fragment) {
+			return true
+		}
+	}
+	return false
+}
+
+func hasCredentialReadVerb(fields []string) bool {
+	for _, field := range fields {
+		switch field {
+		case "cat", "less", "more", "head", "tail", "type", "get-content", "gc", "select-string", "grep", "rg", "sed", "awk":
+			return true
+		case "env", "printenv":
+			return true
+		}
+	}
+	return false
+}
+
+func isRMCommand(field string) bool {
+	return field == "rm" || strings.HasSuffix(field, "/rm")
+}
+
+func isPowerShellDeleteCommand(field string) bool {
+	switch field {
+	case "remove-item", "del", "erase", "rd", "rmdir":
+		return true
+	default:
+		return false
+	}
+}
+
+func isUnsafeDeleteTarget(target string) bool {
+	if target == "." || target == "./" || target == "*" {
+		return true
+	}
+	if target == "/*" {
+		return true
+	}
+	target = strings.TrimSuffix(target, "/*")
+	for _, prefix := range []string{"~/", "$home/", "${home}/", "$env:home/", "$env:userprofile/", "%userprofile%/"} {
+		if strings.HasPrefix(target, prefix) {
+			return true
+		}
+	}
+	for _, prefix := range []string{"/etc/", "/bin/", "/sbin/", "/usr/", "/var/", "/lib/", "/library/", "/system/", "/applications/", "c:/windows/", "c:/program files/"} {
+		if strings.HasPrefix(target, prefix) {
+			return true
+		}
+	}
+	for _, exact := range []string{"/", "~", "$home", "${home}", "$env:home", "$env:userprofile", "%userprofile%", "c:", "c:/", "/etc", "/bin", "/sbin", "/usr", "/var", "/lib", "/library", "/system", "/applications", "c:/windows", "c:/program files"} {
+		if target == exact {
+			return true
+		}
+	}
+	return false
+}
+
+func shellSafetyFields(command string) []string {
+	return strings.Fields(shellSafetyText(command))
+}
+
+func shellSafetyText(command string) string {
+	command = strings.ToLower(command)
+	return strings.NewReplacer(
+		"\\", "/",
+		"\n", " ",
+		"\t", " ",
+		";", " ",
+		"&", " ",
+		"|", " ",
+		"(", " ",
+		")", " ",
+		"\"", "",
+		"'", "",
+		"`", "",
+	).Replace(command)
+}
+
+func readFinalWorkingDir(path string) string {
+	content, err := os.ReadFile(path)
+	if err != nil {
+		return ""
+	}
+	workingDir := strings.TrimPrefix(string(content), "\ufeff")
+	workingDir = strings.TrimSpace(workingDir)
+	if workingDir == "" {
+		return ""
+	}
+	workingDir = normalizeBashWorkingDir(workingDir)
+	info, err := os.Stat(workingDir)
+	if err != nil || !info.IsDir() {
+		return ""
+	}
+	return workingDir
+}
+
+func normalizeBashWorkingDir(workingDir string) string {
+	if runtime.GOOS == "windows" && len(workingDir) >= 3 && workingDir[0] == '/' && workingDir[2] == '/' && isASCIIAlpha(workingDir[1]) {
+		workingDir = strings.ToUpper(string(workingDir[1])) + ":" + workingDir[2:]
+	}
+	workingDir = filepath.Clean(filepath.FromSlash(workingDir))
+	if runtime.GOOS == "windows" && len(workingDir) >= 2 && workingDir[1] == ':' && isASCIIAlpha(workingDir[0]) {
+		workingDir = strings.ToUpper(string(workingDir[0])) + workingDir[1:]
+	}
+	return workingDir
+}
+
+func isASCIIAlpha(b byte) bool {
+	return (b >= 'a' && b <= 'z') || (b >= 'A' && b <= 'Z')
+}
+
+type boundedOutput struct {
+	Limit   int
+	buf     []byte
+	omitted int
+}
+
+func (b *boundedOutput) Write(p []byte) (int, error) {
+	if b.Limit <= 0 {
+		b.omitted += len(p)
+		return len(p), nil
+	}
+	remaining := b.Limit - len(b.buf)
+	if remaining <= 0 {
+		b.omitted += len(p)
+		return len(p), nil
+	}
+	if len(p) <= remaining {
+		b.buf = append(b.buf, p...)
+		return len(p), nil
+	}
+	writeLen := utf8SafePrefixLen(p[:remaining])
+	b.buf = append(b.buf, p[:writeLen]...)
+	b.omitted += len(p) - writeLen
+	return len(p), nil
+}
+
+func (b *boundedOutput) Len() int {
+	return len(b.buf) + b.omitted
+}
+
+func (b *boundedOutput) String(label string) string {
+	safeLen := utf8SafePrefixLen(b.buf)
+	content := string(b.buf[:safeLen])
+	omitted := b.omitted + len(b.buf) - safeLen
+	if omitted == 0 {
+		return content
+	}
+	return content + fmt.Sprintf("\n\n[%s truncated: omitted ~%d tokens]", label, approximateTokensFromBytes(omitted))
+}
+
+func utf8SafePrefixLen(p []byte) int {
+	if len(p) == 0 {
+		return 0
+	}
+	for i := 0; i < len(p); {
+		r, size := utf8.DecodeRune(p[i:])
+		if r == utf8.RuneError && size == 1 {
+			return i
+		}
+		i += size
+	}
+	return len(p)
+}
+
+func approximateTokensFromBytes(n int) int {
+	if n <= 0 {
+		return 0
+	}
+	return max(1, (n+3)/4)
+}
@@ -0,0 +1,258 @@
+package tools
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"unicode/utf8"
+
+	"github.com/ollama/ollama/agent"
+)
+
+func TestBashReportsFinalWorkingDir(t *testing.T) {
+	root := t.TempDir()
+	subdir := filepath.Join(root, "sub")
+	if err := os.Mkdir(subdir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+
+	result, err := (&Bash{}).Execute(context.Background(), agent.ToolContext{WorkingDir: root}, map[string]any{
+		"command": shellTestCommand("cd sub && pwd", "Set-Location sub; Get-Location"),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	wantDir, err := filepath.EvalSymlinks(subdir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.WorkingDir != wantDir {
+		t.Fatalf("working dir = %q, want %q", result.WorkingDir, wantDir)
+	}
+	if !strings.Contains(result.Content, "sub") {
+		t.Fatalf("content = %q, want pwd output", result.Content)
+	}
+}
+
+func TestBashBoundsOutputWhileRunning(t *testing.T) {
+	result, err := (&Bash{}).Execute(context.Background(), agent.ToolContext{WorkingDir: t.TempDir()}, map[string]any{
+		"command": shellTestCommand("yes x | head -c 70000", "[Console]::Out.Write(('x' * 70000))"),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(result.Content, "[stdout truncated: omitted ~") || !strings.Contains(result.Content, " tokens]") {
+		t.Fatalf("content = %q, want stdout truncation marker", result.Content)
+	}
+	if count, want := strings.Count(result.Content, "x"), shellTestCapturedXCount(); count != want {
+		t.Fatalf("captured x count = %d, want %d", count, want)
+	}
+	if len(result.Content) > maxBashOutputBytes+200 {
+		t.Fatalf("content length = %d, want bounded output", len(result.Content))
+	}
+}
+
+func TestBoundedOutputTruncatesAtUTF8Boundary(t *testing.T) {
+	var out boundedOutput
+	out.Limit = len([]byte("abc")) + 1
+
+	if _, err := out.Write([]byte("abcédef")); err != nil {
+		t.Fatal(err)
+	}
+	content := out.String("stdout")
+	if !utf8.ValidString(content) {
+		t.Fatalf("content is not valid UTF-8: %q", content)
+	}
+	if strings.ContainsRune(content, utf8.RuneError) {
+		t.Fatalf("content contains replacement rune: %q", content)
+	}
+	if !strings.HasPrefix(content, "abc\n\n[stdout truncated:") {
+		t.Fatalf("content = %q, want complete ASCII prefix and truncation marker", content)
+	}
+}
+
+func TestBoundedOutputKeepsCompleteUTF8AtBoundary(t *testing.T) {
+	var out boundedOutput
+	out.Limit = len([]byte("abcé"))
+
+	if _, err := out.Write([]byte("abcédef")); err != nil {
+		t.Fatal(err)
+	}
+	if content := out.String("stdout"); !strings.HasPrefix(content, "abcé\n\n[stdout truncated:") {
+		t.Fatalf("content = %q, want complete UTF-8 prefix", content)
+	}
+}
+
+func TestBoundedOutputTrimsTrailingPartialUTF8(t *testing.T) {
+	var out boundedOutput
+	out.Limit = 4
+
+	if _, err := out.Write([]byte{'a', 'b', 'c', 0xc3}); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := out.Write([]byte{0xa9}); err != nil {
+		t.Fatal(err)
+	}
+	if content := out.String("stdout"); !utf8.ValidString(content) || !strings.HasPrefix(content, "abc\n\n[stdout truncated:") {
+		t.Fatalf("content = %q, want valid UTF-8 with partial suffix trimmed", content)
+	}
+}
+
+func TestUTF8SafePrefixRejectsMalformedLeadByte(t *testing.T) {
+	input := []byte{'a', 0xc0, 0x80, 'b'}
+	if got := utf8SafePrefixLen(input); got != 1 {
+		t.Fatalf("safe prefix length = %d, want 1", got)
+	}
+}
+
+func TestBoundedOutputDropsMalformedUTF8(t *testing.T) {
+	var out boundedOutput
+	out.Limit = 4
+
+	if _, err := out.Write([]byte{'a', 0xc0, 0x80, 'b'}); err != nil {
+		t.Fatal(err)
+	}
+	content := out.String("stdout")
+	if !utf8.ValidString(content) {
+		t.Fatalf("content is not valid UTF-8: %q", content)
+	}
+	if strings.ContainsRune(content, utf8.RuneError) {
+		t.Fatalf("content contains replacement rune: %q", content)
+	}
+	if !strings.HasPrefix(content, "a\n\n[stdout truncated:") {
+		t.Fatalf("content = %q, want valid prefix and truncation marker", content)
+	}
+}
+
+func TestBashReportsCanceledCommand(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	result, err := (&Bash{}).Execute(ctx, agent.ToolContext{WorkingDir: t.TempDir()}, map[string]any{
+		"command": shellTestCommand("sleep 10", "Start-Sleep -Seconds 10"),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(result.Content, "Error: command was canceled") {
+		t.Fatalf("content = %q, want canceled message", result.Content)
+	}
+	if strings.Contains(result.Content, "Exit code: -1") {
+		t.Fatalf("content = %q, should not mask cancellation as exit code", result.Content)
+	}
+}
+
+func TestRejectUnsafeShellCommand(t *testing.T) {
+	tests := []struct {
+		name    string
+		command string
+		wantErr bool
+	}{
+		{name: "rm root", command: "rm -rf /", wantErr: true},
+		{name: "sudo rm root", command: "sudo rm -rf -- /", wantErr: true},
+		{name: "rm home", command: "rm -fr $HOME", wantErr: true},
+		{name: "rm root wildcard", command: "rm -rf /*", wantErr: true},
+		{name: "rm system subdir", command: "rm -rf /etc/ssh", wantErr: true},
+		{name: "rm cwd", command: "rm -rf .", wantErr: true},
+		{name: "powershell remove root", command: `Remove-Item -Recurse -Force C:\`, wantErr: true},
+		{name: "powershell remove system subdir", command: `Remove-Item -Recurse -Force C:\Windows\Temp`, wantErr: true},
+		{name: "ssh private key", command: "cat ~/.ssh/id_rsa", wantErr: true},
+		{name: "aws credentials", command: "Get-Content $HOME/.aws/credentials", wantErr: true},
+		{name: "shadow", command: "head /etc/shadow", wantErr: true},
+		{name: "netrc", command: "cat ~/.netrc", wantErr: true},
+		{name: "docker config", command: "cat ~/.docker/config.json", wantErr: true},
+		{name: "gnupg dir", command: "cat ~/.gnupg/private-keys-v1.d/key", wantErr: true},
+		{name: "gh hosts", command: "cat ~/.config/gh/hosts.yml", wantErr: true},
+		{name: "ssh config", command: "cat ~/.ssh/config", wantErr: true},
+		{name: "printenv dump", command: "printenv", wantErr: false},
+		{name: "delete build dir", command: "rm -rf build", wantErr: false},
+		{name: "read project file", command: "cat README.md", wantErr: false},
+		{name: "mention key text", command: "rg id_rsa docs", wantErr: false},
+		{name: "env example", command: "cat .env.example", wantErr: false},
+		{name: "rm build then unrelated tilde path", command: "rm -rf build && echo ~/.ssh/config", wantErr: false},
+		{name: "rm build then unrelated slash path", command: "rm -rf build; cat /etc/passwd", wantErr: false},
+		{name: "rm build then unrelated star glob", command: "rm -rf build && ls *.go", wantErr: false},
+		{name: "rm multiple targets one unsafe", command: "rm -rf build /etc", wantErr: true},
+		{name: "rm unsafe then safe piped", command: "rm -rf / | tee log", wantErr: true},
+		{name: "rm unsafe via command substitution", command: "rm -rf $(echo /)", wantErr: true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			err := rejectUnsafeShellCommand(tt.command)
+			if tt.wantErr && err == nil {
+				t.Fatal("expected unsafe command to be rejected")
+			}
+			if !tt.wantErr && err != nil {
+				t.Fatalf("command rejected: %v", err)
+			}
+		})
+	}
+}
+
+func TestBashRejectsUnsafeCommandBeforeExecution(t *testing.T) {
+	_, err := (&Bash{}).Execute(context.Background(), agent.ToolContext{WorkingDir: t.TempDir()}, map[string]any{
+		"command": "rm -rf /",
+	})
+	if err == nil || !strings.Contains(err.Error(), "refusing to run unsafe command") {
+		t.Fatalf("err = %v, want unsafe command rejection", err)
+	}
+}
+
+func shellTestCommand(unix, windows string) string {
+	if runtime.GOOS == "windows" {
+		return windows
+	}
+	return unix
+}
+
+func shellTestCapturedXCount() int {
+	if runtime.GOOS == "windows" {
+		return maxBashOutputBytes
+	}
+	return maxBashOutputBytes / 2
+}
+
+func TestReadFinalWorkingDirRejectsInvalidPaths(t *testing.T) {
+	dir := t.TempDir()
+	cwdFile := filepath.Join(dir, "cwd")
+	notDir := filepath.Join(dir, "file.txt")
+	if err := os.WriteFile(notDir, []byte("not a dir"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := os.WriteFile(cwdFile, []byte(notDir+"\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if got := readFinalWorkingDir(cwdFile); got != "" {
+		t.Fatalf("regular file cwd = %q, want empty", got)
+	}
+
+	if err := os.WriteFile(cwdFile, []byte(filepath.Join(dir, "missing")+"\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if got := readFinalWorkingDir(cwdFile); got != "" {
+		t.Fatalf("missing cwd = %q, want empty", got)
+	}
+
+	if err := os.WriteFile(cwdFile, []byte(dir+"\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if got := readFinalWorkingDir(cwdFile); got != dir {
+		t.Fatalf("directory cwd = %q, want %q", got, dir)
+	}
+}
+
+func TestNormalizeBashWorkingDirWindowsDriveLetter(t *testing.T) {
+	if runtime.GOOS != "windows" {
+		t.Skip("windows path normalization")
+	}
+	got := normalizeBashWorkingDir("/c/Users/jdoe/project")
+	want := filepath.Clean(`C:\Users\jdoe\project`)
+	if got != want {
+		t.Fatalf("working dir = %q, want %q", got, want)
+	}
+}
@@ -0,0 +1,49 @@
+//go:build !windows
+
+package tools
+
+import (
+	"context"
+	"os/exec"
+	"strings"
+	"syscall"
+)
+
+func shellToolName() string {
+	return "bash"
+}
+
+func shellToolDescription() string {
+	return "Execute a bash command on the system. Use this to inspect files, run tests, and perform development tasks."
+}
+
+func shellCommandDescription() string {
+	return "The bash command to execute."
+}
+
+func newBashCommand(ctx context.Context, command, cwdPath string) *exec.Cmd {
+	script := command + "\n__ollama_status=$?\npwd -P > " + shellQuote(cwdPath) + "\nexit $__ollama_status"
+	cmd := exec.CommandContext(ctx, "bash", "-c", script)
+	configureBashCommand(cmd)
+	return cmd
+}
+
+func shellQuote(value string) string {
+	return "'" + strings.ReplaceAll(value, "'", "'\\''") + "'"
+}
+
+func configureBashCommand(cmd *exec.Cmd) {
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+}
+
+func runBashCommand(cmd *exec.Cmd) error {
+	return cmd.Run()
+}
+
+func killBashCommand(cmd *exec.Cmd) error {
+	if cmd == nil || cmd.Process == nil {
+		return nil
+	}
+	_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
+	return nil
+}
@@ -0,0 +1,40 @@
+//go:build !windows
+
+package tools
+
+import (
+	"context"
+	"os/exec"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/agent"
+)
+
+func TestConfigureBashCommandSetsProcessGroup(t *testing.T) {
+	cmd := exec.Command("bash", "-c", "true")
+	configureBashCommand(cmd)
+	if cmd.SysProcAttr == nil || !cmd.SysProcAttr.Setpgid {
+		t.Fatalf("configureBashCommand should start bash in a new process group")
+	}
+}
+
+func TestBashWaitDelayBoundsBackgroundOutputPipe(t *testing.T) {
+	start := time.Now()
+	result, err := (&Bash{}).Execute(context.Background(), agent.ToolContext{WorkingDir: t.TempDir()}, map[string]any{
+		"command": "sleep 5 & echo done",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if elapsed := time.Since(start); elapsed > bashWaitDelay+2*time.Second {
+		t.Fatalf("command elapsed = %s, want bounded near %s", elapsed, bashWaitDelay)
+	}
+	if !strings.Contains(result.Content, "done") {
+		t.Fatalf("content = %q, want command output", result.Content)
+	}
+	if !strings.Contains(result.Content, "output pipes did not close") {
+		t.Fatalf("content = %q, want wait delay message", result.Content)
+	}
+}
@@ -0,0 +1,134 @@
+//go:build windows
+
+package tools
+
+import (
+	"context"
+	"os/exec"
+	"strings"
+	"sync"
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+var bashJobHandles sync.Map
+
+func shellToolName() string {
+	return "powershell"
+}
+
+func shellToolDescription() string {
+	return "Execute a PowerShell command on the system. Use this to inspect files, run tests, and perform development tasks."
+}
+
+func shellCommandDescription() string {
+	return "The PowerShell command to execute."
+}
+
+func newBashCommand(ctx context.Context, command, cwdPath string) *exec.Cmd {
+	return exec.CommandContext(
+		ctx,
+		"powershell.exe",
+		"-NoLogo",
+		"-NoProfile",
+		"-NonInteractive",
+		"-ExecutionPolicy",
+		"Bypass",
+		"-Command",
+		powerShellCommandScript(command, cwdPath),
+	)
+}
+
+func powerShellCommandScript(command, cwdPath string) string {
+	cwdPath = powerShellSingleQuote(cwdPath)
+	return strings.Join([]string{
+		"$__ollama_status = 0",
+		". {",
+		"try {",
+		command,
+		"  $__ollama_success = $?",
+		"  $__ollama_last_exit = $global:LASTEXITCODE",
+		"  if ($__ollama_success) {",
+		"    $__ollama_status = 0",
+		"  } elseif ($__ollama_last_exit -is [int] -and $__ollama_last_exit -ne 0) {",
+		"    $__ollama_status = $__ollama_last_exit",
+		"  } else {",
+		"    $__ollama_status = 1",
+		"  }",
+		"} catch {",
+		"  Write-Error $_",
+		"  $__ollama_status = 1",
+		"} finally {",
+		"  try { [System.IO.File]::WriteAllText(" + cwdPath + ", (Get-Location).ProviderPath, [System.Text.Encoding]::UTF8) } catch {}",
+		"}",
+		"} | Out-String -Stream -Width 4096",
+		"exit $__ollama_status",
+	}, "\n")
+}
+
+func powerShellSingleQuote(value string) string {
+	return "'" + strings.ReplaceAll(value, "'", "''") + "'"
+}
+
+func runBashCommand(cmd *exec.Cmd) error {
+	if err := cmd.Start(); err != nil {
+		return err
+	}
+	if job, err := createBashJob(cmd.Process.Pid); err == nil {
+		bashJobHandles.Store(cmd.Process.Pid, job)
+		defer releaseBashJob(cmd.Process.Pid)
+	}
+	return cmd.Wait()
+}
+
+func killBashCommand(cmd *exec.Cmd) error {
+	if cmd == nil || cmd.Process == nil {
+		return nil
+	}
+	releaseBashJob(cmd.Process.Pid)
+	_ = cmd.Process.Kill()
+	return nil
+}
+
+func createBashJob(pid int) (windows.Handle, error) {
+	job, err := windows.CreateJobObject(nil, nil)
+	if err != nil {
+		return 0, err
+	}
+
+	info := windows.JOBOBJECT_EXTENDED_LIMIT_INFORMATION{}
+	info.BasicLimitInformation.LimitFlags = windows.JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE
+	if _, err := windows.SetInformationJobObject(
+		job,
+		windows.JobObjectExtendedLimitInformation,
+		uintptr(unsafe.Pointer(&info)),
+		uint32(unsafe.Sizeof(info)),
+	); err != nil {
+		_ = windows.CloseHandle(job)
+		return 0, err
+	}
+
+	process, err := windows.OpenProcess(windows.PROCESS_SET_QUOTA|windows.PROCESS_TERMINATE, false, uint32(pid))
+	if err != nil {
+		_ = windows.CloseHandle(job)
+		return 0, err
+	}
+	defer windows.CloseHandle(process)
+
+	if err := windows.AssignProcessToJobObject(job, process); err != nil {
+		_ = windows.CloseHandle(job)
+		return 0, err
+	}
+	return job, nil
+}
+
+func releaseBashJob(pid int) {
+	value, ok := bashJobHandles.LoadAndDelete(pid)
+	if !ok {
+		return
+	}
+	if job, ok := value.(windows.Handle); ok {
+		_ = windows.CloseHandle(job)
+	}
+}
@@ -0,0 +1,15 @@
+//go:build windows
+
+package tools
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestPowerShellCommandScriptUsesWideOutString(t *testing.T) {
+	script := powerShellCommandScript("Get-ChildItem", `C:\cwd.txt`)
+	if !strings.Contains(script, "Out-String -Stream -Width 4096") {
+		t.Fatalf("script = %q, want explicit Out-String width", script)
+	}
+}
@@ -0,0 +1,556 @@
+package tools
+
+import (
+	"bufio"
+	"context"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+
+	"github.com/ollama/ollama/agent"
+	"github.com/ollama/ollama/api"
+)
+
+const (
+	maxReadBytes = 200000
+)
+
+type Read struct{}
+
+func (r *Read) Name() string {
+	return "read"
+}
+
+func (r *Read) Description() string {
+	return "Read a text file from the current working directory."
+}
+
+func (r *Read) Schema() api.ToolFunction {
+	props := api.NewToolPropertiesMap()
+	props.Set("path", api.ToolProperty{
+		Type:        api.PropertyType{"string"},
+		Description: "Path to the file to read, relative to the working directory.",
+	})
+	props.Set("start", api.ToolProperty{
+		Type:        api.PropertyType{"integer"},
+		Description: "Optional 1-based line to start reading from.",
+	})
+	props.Set("end", api.ToolProperty{
+		Type:        api.PropertyType{"integer"},
+		Description: "Optional 1-based inclusive line to stop reading at.",
+	})
+	return api.ToolFunction{
+		Name:        r.Name(),
+		Description: r.Description(),
+		Parameters: api.ToolFunctionParameters{
+			Type:       "object",
+			Properties: props,
+			Required:   []string{"path"},
+		},
+	}
+}
+
+func (r *Read) RequiresApproval(map[string]any) bool {
+	return true
+}
+
+func (r *Read) Execute(ctx context.Context, toolCtx agent.ToolContext, args map[string]any) (agent.ToolResult, error) {
+	path, ok := args["path"].(string)
+	if !ok || strings.TrimSpace(path) == "" {
+		return agent.ToolResult{}, fmt.Errorf("path parameter is required")
+	}
+
+	file, info, err := openRegularFile(toolCtx.WorkingDir, path, true)
+	if err != nil {
+		return agent.ToolResult{}, err
+	}
+	defer file.Close()
+
+	selection, err := readSelectionFromArgs(args)
+	if err != nil {
+		return agent.ToolResult{}, err
+	}
+	if !selection.enabled && info.Size() > maxReadBytes {
+		return agent.ToolResult{}, fmt.Errorf("%s is too large to read (%d bytes)", path, info.Size())
+	}
+
+	select {
+	case <-ctx.Done():
+		return agent.ToolResult{}, ctx.Err()
+	default:
+	}
+
+	var content string
+	if selection.enabled {
+		content, err = readLineSelection(file, selection)
+	} else {
+		var contentBytes []byte
+		contentBytes, err = readAllWithinLimit(file, maxReadBytes)
+		content = string(contentBytes)
+	}
+	if err != nil {
+		return agent.ToolResult{}, err
+	}
+	return agent.ToolResult{Content: content}, nil
+}
+
+type Edit struct{}
+
+func (e *Edit) Name() string {
+	return "edit"
+}
+
+func (e *Edit) Description() string {
+	return "Edit a text file in the current working directory by replacing exact text."
+}
+
+func (e *Edit) Schema() api.ToolFunction {
+	props := api.NewToolPropertiesMap()
+	props.Set("path", api.ToolProperty{
+		Type:        api.PropertyType{"string"},
+		Description: "Path to the file to edit, relative to the working directory.",
+	})
+	props.Set("old_text", api.ToolProperty{
+		Type:        api.PropertyType{"string"},
+		Description: "Exact text to replace.",
+	})
+	props.Set("new_text", api.ToolProperty{
+		Type:        api.PropertyType{"string"},
+		Description: "Replacement text.",
+	})
+	props.Set("replace_all", api.ToolProperty{
+		Type:        api.PropertyType{"boolean"},
+		Description: "Replace every occurrence. Defaults to false and requires old_text to match exactly once.",
+	})
+	return api.ToolFunction{
+		Name:        e.Name(),
+		Description: e.Description(),
+		Parameters: api.ToolFunctionParameters{
+			Type:       "object",
+			Properties: props,
+			Required:   []string{"path", "old_text", "new_text"},
+		},
+	}
+}
+
+func (e *Edit) RequiresApproval(map[string]any) bool {
+	return true
+}
+
+func (e *Edit) Execute(ctx context.Context, toolCtx agent.ToolContext, args map[string]any) (agent.ToolResult, error) {
+	path, ok := args["path"].(string)
+	if !ok || strings.TrimSpace(path) == "" {
+		return agent.ToolResult{}, fmt.Errorf("path parameter is required")
+	}
+
+	oldText, ok := args["old_text"].(string)
+	if !ok || oldText == "" {
+		return agent.ToolResult{}, fmt.Errorf("old_text parameter is required")
+	}
+
+	newText, ok := args["new_text"].(string)
+	if !ok {
+		return agent.ToolResult{}, fmt.Errorf("new_text parameter is required")
+	}
+
+	replaceAll, _ := args["replace_all"].(bool)
+
+	if err := rejectFinalSymlink(toolCtx.WorkingDir, path); err != nil {
+		return agent.ToolResult{}, err
+	}
+
+	file, info, err := openRegularFile(toolCtx.WorkingDir, path, false)
+	if err != nil {
+		return agent.ToolResult{}, err
+	}
+	if info.Size() > maxReadBytes {
+		file.Close()
+		return agent.ToolResult{}, fmt.Errorf("%s is too large to edit (%d bytes)", path, info.Size())
+	}
+
+	select {
+	case <-ctx.Done():
+		file.Close()
+		return agent.ToolResult{}, ctx.Err()
+	default:
+	}
+
+	contentBytes, err := readAllWithinLimit(file, maxReadBytes)
+	if closeErr := file.Close(); err == nil && closeErr != nil {
+		err = closeErr
+	}
+	if err != nil {
+		return agent.ToolResult{}, err
+	}
+	content := string(contentBytes)
+	matches := strings.Count(content, oldText)
+	if matches == 0 {
+		return agent.ToolResult{}, fmt.Errorf("old_text was not found in %s", path)
+	}
+	if matches > 1 && !replaceAll {
+		return agent.ToolResult{}, fmt.Errorf("old_text matched %d times in %s; set replace_all to true to replace every match", matches, path)
+	}
+
+	var updated string
+	if replaceAll {
+		updated = strings.ReplaceAll(content, oldText, newText)
+	} else {
+		updated = strings.Replace(content, oldText, newText, 1)
+	}
+	if len(updated) > maxReadBytes {
+		return agent.ToolResult{}, fmt.Errorf("edited content is too large (%d bytes)", len(updated))
+	}
+
+	if err := writeFileAtomic(toolCtx.WorkingDir, path, []byte(updated), info.Mode().Perm()); err != nil {
+		return agent.ToolResult{}, err
+	}
+
+	return agent.ToolResult{Content: fmt.Sprintf("Updated %s (%d replacement%s).", path, matches, plural(matches))}, nil
+}
+
+func cleanRelativePath(path string) (string, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return "", fmt.Errorf("path parameter is required")
+	}
+	if filepath.IsAbs(path) {
+		return "", fmt.Errorf("absolute paths are not allowed")
+	}
+	cleaned := filepath.Clean(path)
+	if cleaned == "." || cleaned == ".." || strings.HasPrefix(cleaned, ".."+string(os.PathSeparator)) {
+		return "", fmt.Errorf("path escapes working directory")
+	}
+	return cleaned, nil
+}
+
+func openRegularFile(workingDir, path string, allowAbsolute bool) (*os.File, os.FileInfo, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return nil, nil, fmt.Errorf("path parameter is required")
+	}
+	if allowAbsolute && filepath.IsAbs(path) {
+		cleaned := filepath.Clean(path)
+		info, err := os.Lstat(cleaned)
+		if err != nil {
+			return nil, nil, err
+		}
+		if info.Mode()&os.ModeSymlink != 0 {
+			return nil, nil, fmt.Errorf("%s is a symlink; read the target file directly", path)
+		}
+		if err := rejectNonRegularFile(path, info); err != nil {
+			return nil, nil, err
+		}
+		file, err := os.Open(cleaned)
+		if err != nil {
+			return nil, nil, err
+		}
+		info, err = file.Stat()
+		if err != nil {
+			file.Close()
+			return nil, nil, err
+		}
+		if err := rejectNonRegularFile(path, info); err != nil {
+			file.Close()
+			return nil, nil, err
+		}
+		return file, info, nil
+	}
+
+	rel, err := cleanRelativePath(path)
+	if err != nil {
+		return nil, nil, err
+	}
+	root, err := openWorkingRoot(workingDir)
+	if err != nil {
+		return nil, nil, err
+	}
+	defer root.Close()
+
+	if _, err := regularRootFileInfo(root, rel, path); err != nil {
+		return nil, nil, err
+	}
+	file, err := root.Open(rel)
+	if err != nil {
+		return nil, nil, rootPathError(err)
+	}
+	info, err := file.Stat()
+	if err != nil {
+		file.Close()
+		return nil, nil, err
+	}
+	if err := rejectNonRegularFile(path, info); err != nil {
+		file.Close()
+		return nil, nil, err
+	}
+	return file, info, nil
+}
+
+func regularRootFileInfo(root *os.Root, rel, path string) (os.FileInfo, error) {
+	info, err := root.Lstat(rel)
+	if err != nil {
+		return nil, rootPathError(err)
+	}
+	// Reject symlinks outright. os.Root.Open follows symlinks via openat
+	// without O_NOFOLLOW, so a symlink inside the working root that points
+	// outside it (e.g. ./notes -> ~/.ssh/id_rsa) would otherwise be read
+	// transparently, bypassing the working-directory confinement that the
+	// bash denylist enforces for direct credential reads. The caller must
+	// operate on the real target file instead.
+	if info.Mode()&os.ModeSymlink != 0 {
+		return nil, fmt.Errorf("%s is a symlink; read the target file directly", path)
+	}
+	if err := rejectNonRegularFile(path, info); err != nil {
+		return nil, err
+	}
+	return info, nil
+}
+
+func rejectNonRegularFile(path string, info os.FileInfo) error {
+	if info.IsDir() {
+		return fmt.Errorf("%s is a directory", path)
+	}
+	if !info.Mode().IsRegular() {
+		return fmt.Errorf("%s is not a regular file", path)
+	}
+	return nil
+}
+
+func writeFileAtomic(workingDir, path string, data []byte, perm os.FileMode) error {
+	rel, err := cleanRelativePath(path)
+	if err != nil {
+		return err
+	}
+	root, err := openWorkingRoot(workingDir)
+	if err != nil {
+		return err
+	}
+	defer root.Close()
+	if err := rejectRootFinalSymlink(root, rel, path); err != nil {
+		return err
+	}
+
+	parent, name := filepath.Split(rel)
+	tmpBase := fmt.Sprintf(".%s.ollama-tmp-%d", name, os.Getpid())
+	for i := 0; ; i++ {
+		candidateName := tmpBase
+		if i > 0 {
+			candidateName = fmt.Sprintf("%s-%d", tmpBase, i)
+		}
+		candidate := filepath.Join(parent, candidateName)
+		file, err := root.OpenFile(candidate, os.O_WRONLY|os.O_CREATE|os.O_EXCL, perm)
+		if os.IsExist(err) {
+			continue
+		}
+		if err != nil {
+			return rootPathError(err)
+		}
+		if err := file.Chmod(perm); err != nil {
+			closeErr := file.Close()
+			_ = root.Remove(candidate)
+			if closeErr != nil {
+				return closeErr
+			}
+			return err
+		}
+		writeErr := writeAllAndSync(file, data)
+		closeErr := file.Close()
+		if writeErr != nil || closeErr != nil {
+			_ = root.Remove(candidate)
+			if writeErr != nil {
+				return writeErr
+			}
+			return closeErr
+		}
+		if err := root.Rename(candidate, rel); err != nil {
+			_ = root.Remove(candidate)
+			return rootPathError(err)
+		}
+		return nil
+	}
+}
+
+func rejectFinalSymlink(workingDir, path string) error {
+	rel, err := cleanRelativePath(path)
+	if err != nil {
+		return err
+	}
+	root, err := openWorkingRoot(workingDir)
+	if err != nil {
+		return err
+	}
+	defer root.Close()
+	return rejectRootFinalSymlink(root, rel, path)
+}
+
+func rejectRootFinalSymlink(root *os.Root, rel, path string) error {
+	info, err := root.Lstat(rel)
+	if err != nil {
+		return rootPathError(err)
+	}
+	if info.Mode()&os.ModeSymlink != 0 {
+		return fmt.Errorf("%s is a symlink; edit the target file directly", path)
+	}
+	return nil
+}
+
+func rootPathError(err error) error {
+	if err != nil && strings.Contains(err.Error(), "path escapes") {
+		return fmt.Errorf("path escapes working directory")
+	}
+	return err
+}
+
+func openWorkingRoot(workingDir string) (*os.Root, error) {
+	base, err := workingDirAbs(workingDir)
+	if err != nil {
+		return nil, err
+	}
+	return os.OpenRoot(base)
+}
+
+func writeAllAndSync(file *os.File, data []byte) error {
+	if _, err := file.Write(data); err != nil {
+		return err
+	}
+	return file.Sync()
+}
+
+func readAllWithinLimit(reader io.Reader, limit int) ([]byte, error) {
+	if limit < 0 {
+		limit = 0
+	}
+	content, err := io.ReadAll(io.LimitReader(reader, int64(limit)+1))
+	if err != nil {
+		return nil, err
+	}
+	if len(content) > limit {
+		return nil, fmt.Errorf("content is too large (%d byte limit)", limit)
+	}
+	return content, nil
+}
+
+func workingDirAbs(workingDir string) (string, error) {
+	base := workingDir
+	if base == "" {
+		var err error
+		base, err = os.Getwd()
+		if err != nil {
+			return "", err
+		}
+	}
+	return canonicalPath(base)
+}
+
+func canonicalPath(path string) (string, error) {
+	abs, err := filepath.Abs(path)
+	if err != nil {
+		return "", err
+	}
+	resolved, err := filepath.EvalSymlinks(abs)
+	if err == nil {
+		return resolved, nil
+	}
+	return abs, nil
+}
+
+type readSelection struct {
+	enabled bool
+	start   int
+	end     int
+}
+
+func readSelectionFromArgs(args map[string]any) (readSelection, error) {
+	selection := readSelection{start: 1}
+
+	if start, ok, err := intReadArg(args, "start"); err != nil {
+		return readSelection{}, err
+	} else if ok {
+		selection.enabled = true
+		selection.start = start
+	}
+	if end, ok, err := intReadArg(args, "end"); err != nil {
+		return readSelection{}, err
+	} else if ok {
+		selection.enabled = true
+		selection.end = end
+	}
+
+	if !selection.enabled {
+		return selection, nil
+	}
+	if selection.start < 1 {
+		return readSelection{}, fmt.Errorf("start must be greater than 0")
+	}
+	if selection.end > 0 && selection.end < selection.start {
+		return readSelection{}, fmt.Errorf("end must be greater than or equal to start")
+	}
+	return selection, nil
+}
+
+func readLineSelection(file *os.File, selection readSelection) (string, error) {
+	reader := bufio.NewReader(file)
+	var b strings.Builder
+	for lineNo := 1; ; {
+		line, err := reader.ReadSlice('\n')
+		if lineNo >= selection.start && (selection.end == 0 || lineNo <= selection.end) {
+			if b.Len()+len(line) > maxReadBytes {
+				return "", fmt.Errorf("selected content is too large (%d byte limit)", maxReadBytes)
+			}
+			b.Write(line)
+		}
+		if err != nil {
+			if err == bufio.ErrBufferFull {
+				continue
+			}
+			if err == io.EOF {
+				break
+			}
+			return "", err
+		}
+		if selection.end > 0 && lineNo >= selection.end {
+			break
+		}
+		lineNo++
+	}
+	return b.String(), nil
+}
+
+func intReadArg(args map[string]any, key string) (int, bool, error) {
+	value, ok := args[key]
+	if !ok {
+		return 0, false, nil
+	}
+	switch v := value.(type) {
+	case int:
+		return v, true, nil
+	case int64:
+		return int(v), true, nil
+	case float64:
+		if v != float64(int(v)) {
+			return 0, true, fmt.Errorf("%s must be a whole number", key)
+		}
+		return int(v), true, nil
+	case string:
+		v = strings.TrimSpace(v)
+		if v == "" {
+			return 0, false, nil
+		}
+		n, err := strconv.Atoi(v)
+		if err != nil {
+			return 0, true, fmt.Errorf("%s must be a whole number", key)
+		}
+		return n, true, nil
+	default:
+		return 0, true, fmt.Errorf("%s must be a whole number", key)
+	}
+}
+
+func plural(n int) string {
+	if n == 1 {
+		return ""
+	}
+	return "s"
+}
@@ -0,0 +1,338 @@
+package tools
+
+import (
+	"context"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/ollama/ollama/agent"
+)
+
+func TestEditReplacesUniqueText(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "note.txt")
+	if err := os.WriteFile(path, []byte("hello world\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	result, err := (&Edit{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path":     "note.txt",
+		"old_text": "hello",
+		"new_text": "hi",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(result.Content, "Updated note.txt") {
+		t.Fatalf("result = %q", result.Content)
+	}
+
+	content, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(content) != "hi world\n" {
+		t.Fatalf("content = %q", content)
+	}
+}
+
+func TestEditRequiresUniqueMatchByDefault(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "note.txt")
+	if err := os.WriteFile(path, []byte("same same\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := (&Edit{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path":     "note.txt",
+		"old_text": "same",
+		"new_text": "other",
+	})
+	if err == nil {
+		t.Fatal("expected ambiguous edit to fail")
+	}
+	if !strings.Contains(err.Error(), "matched 2 times") {
+		t.Fatalf("err = %v", err)
+	}
+}
+
+func TestEditRejectsEscapingPath(t *testing.T) {
+	dir := t.TempDir()
+	_, err := (&Edit{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path":     "../outside.txt",
+		"old_text": "old",
+		"new_text": "new",
+	})
+	if err == nil {
+		t.Fatal("expected escaping path to fail")
+	}
+	if !strings.Contains(err.Error(), "path escapes working directory") {
+		t.Fatalf("err = %v", err)
+	}
+}
+
+func TestEditRejectsSymlinkEscape(t *testing.T) {
+	dir := t.TempDir()
+	outside := t.TempDir()
+	if err := os.WriteFile(filepath.Join(outside, "note.txt"), []byte("old\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Symlink(outside, filepath.Join(dir, "link")); err != nil {
+		t.Skipf("symlinks unavailable: %v", err)
+	}
+
+	_, err := (&Edit{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path":     filepath.Join("link", "note.txt"),
+		"old_text": "old",
+		"new_text": "new",
+	})
+	if err == nil {
+		t.Fatal("expected symlink escape to fail")
+	}
+	if !strings.Contains(err.Error(), "path escapes working directory") {
+		t.Fatalf("err = %v", err)
+	}
+
+	content, err := os.ReadFile(filepath.Join(outside, "note.txt"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(content) != "old\n" {
+		t.Fatalf("outside content changed to %q", content)
+	}
+}
+
+func TestEditRejectsFinalSymlink(t *testing.T) {
+	dir := t.TempDir()
+	target := filepath.Join(dir, "target.txt")
+	if err := os.WriteFile(target, []byte("old\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	link := filepath.Join(dir, "link.txt")
+	if err := os.Symlink("target.txt", link); err != nil {
+		t.Skipf("symlinks unavailable: %v", err)
+	}
+
+	_, err := (&Edit{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path":     "link.txt",
+		"old_text": "old",
+		"new_text": "new",
+	})
+	if err == nil {
+		t.Fatal("expected final symlink edit to fail")
+	}
+	if !strings.Contains(err.Error(), "is a symlink") {
+		t.Fatalf("err = %v", err)
+	}
+	content, err := os.ReadFile(target)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(content) != "old\n" {
+		t.Fatalf("target content changed to %q", content)
+	}
+	info, err := os.Lstat(link)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if info.Mode()&os.ModeSymlink == 0 {
+		t.Fatalf("link mode = %v, want symlink", info.Mode())
+	}
+}
+
+func TestReadRejectsParentOutsideCurrentWorkingDir(t *testing.T) {
+	root := t.TempDir()
+	subdir := filepath.Join(root, "sub")
+	if err := os.Mkdir(subdir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(root, "note.txt"), []byte("hello"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: subdir}, map[string]any{
+		"path": "../note.txt",
+	})
+	if err == nil {
+		t.Fatal("expected parent path to fail")
+	}
+	if !strings.Contains(err.Error(), "path escapes working directory") {
+		t.Fatalf("err = %v", err)
+	}
+}
+
+func TestReadRequiresApproval(t *testing.T) {
+	if !agent.ToolRequiresApproval((&Read{}), map[string]any{"path": "note.txt"}) {
+		t.Fatal("read should require approval")
+	}
+}
+
+func TestReadDefaultsToEntireFile(t *testing.T) {
+	dir := t.TempDir()
+	content := "one\ntwo\nthree\n"
+	if err := os.WriteFile(filepath.Join(dir, "note.txt"), []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	result, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path": "note.txt",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Content != content {
+		t.Fatalf("content = %q", result.Content)
+	}
+}
+
+func TestReadAllowsAbsolutePath(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "note.txt")
+	content := "one\ntwo\nthree\n"
+	if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	result, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: t.TempDir()}, map[string]any{
+		"path": path,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Content != content {
+		t.Fatalf("content = %q", result.Content)
+	}
+}
+
+func TestReadRejectsAbsoluteSymlink(t *testing.T) {
+	dir := t.TempDir()
+	target := filepath.Join(dir, "target.txt")
+	if err := os.WriteFile(target, []byte("hello\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	link := filepath.Join(dir, "alias")
+	if err := os.Symlink(target, link); err != nil {
+		t.Skipf("symlinks unavailable: %v", err)
+	}
+
+	_, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: t.TempDir()}, map[string]any{
+		"path": link,
+	})
+	if err == nil {
+		t.Fatal("expected absolute symlink to be rejected")
+	}
+	if !strings.Contains(err.Error(), "symlink") {
+		t.Fatalf("err = %v, want symlink rejection", err)
+	}
+}
+
+func TestReadStartEnd(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, "note.txt"), []byte("one\ntwo\nthree\nfour\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	result, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path":  "note.txt",
+		"start": 2,
+		"end":   3,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Content != "two\nthree\n" {
+		t.Fatalf("content = %q", result.Content)
+	}
+}
+
+func TestReadStartOnly(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, "note.txt"), []byte("one\ntwo\nthree\nfour\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	result, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path":  "note.txt",
+		"start": 3,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Content != "three\nfour\n" {
+		t.Fatalf("content = %q", result.Content)
+	}
+}
+
+func TestReadEndOnly(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, "note.txt"), []byte("one\ntwo\nthree\nfour\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	result, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path": "note.txt",
+		"end":  2,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Content != "one\ntwo\n" {
+		t.Fatalf("content = %q", result.Content)
+	}
+}
+
+func TestReadSelectionRejectsHugeSingleLine(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, "note.txt"), []byte(strings.Repeat("x", maxReadBytes+1)), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path":  "note.txt",
+		"start": 1,
+		"end":   1,
+	})
+	if err == nil {
+		t.Fatal("expected huge selected line to fail")
+	}
+	if !strings.Contains(err.Error(), "selected content is too large") {
+		t.Fatalf("err = %v", err)
+	}
+}
+
+func TestReadAllWithinLimitRejectsGrowingRead(t *testing.T) {
+	reader := io.MultiReader(
+		strings.NewReader(strings.Repeat("x", maxReadBytes)),
+		strings.NewReader("x"),
+	)
+
+	_, err := readAllWithinLimit(reader, maxReadBytes)
+	if err == nil {
+		t.Fatal("expected over-limit read to fail")
+	}
+	if !strings.Contains(err.Error(), "content is too large") {
+		t.Fatalf("err = %v", err)
+	}
+}
+
+func TestReadRejectsInvalidRange(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, "note.txt"), []byte("one\ntwo\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path":  "note.txt",
+		"start": 4,
+		"end":   2,
+	})
+	if err == nil {
+		t.Fatal("expected invalid range to fail")
+	}
+	if !strings.Contains(err.Error(), "end must") {
+		t.Fatalf("err = %v", err)
+	}
+}
@@ -0,0 +1,121 @@
+//go:build !windows
+
+package tools
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"syscall"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/agent"
+)
+
+func TestOpenRegularFileRejectsFIFO(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "pipe")
+	if err := syscall.Mkfifo(path, 0o600); err != nil {
+		t.Skipf("mkfifo unavailable: %v", err)
+	}
+
+	done := make(chan error, 1)
+	go func() {
+		file, _, err := openRegularFile(dir, "pipe", false)
+		if file != nil {
+			file.Close()
+		}
+		done <- err
+	}()
+
+	select {
+	case err := <-done:
+		if err == nil {
+			t.Fatal("expected FIFO to be rejected")
+		}
+		if !strings.Contains(err.Error(), "not a regular file") {
+			t.Fatalf("err = %v", err)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("openRegularFile blocked on FIFO")
+	}
+}
+
+func TestEditPreservesModeDespiteUmask(t *testing.T) {
+	oldUmask := syscall.Umask(0o077)
+	defer syscall.Umask(oldUmask)
+
+	dir := t.TempDir()
+	path := filepath.Join(dir, "note.txt")
+	if err := os.WriteFile(path, []byte("hello\n"), 0o666); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Chmod(path, 0o666); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := (&Edit{}).Execute(context.Background(), agent.ToolContext{WorkingDir: dir}, map[string]any{
+		"path":     "note.txt",
+		"old_text": "hello",
+		"new_text": "hi",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	info, err := os.Stat(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := info.Mode().Perm(); got != 0o666 {
+		t.Fatalf("mode = %#o, want 0666", got)
+	}
+}
+
+func TestReadRejectsSymlinkEscapingWorkingDir(t *testing.T) {
+	root := t.TempDir()
+	secret := filepath.Join(t.TempDir(), "secret.txt")
+	if err := os.WriteFile(secret, []byte("top secret\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	link := filepath.Join(root, "notes")
+	if err := os.Symlink(secret, link); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: root}, map[string]any{
+		"path": "notes",
+	})
+	if err == nil {
+		t.Fatal("expected symlink escaping working dir to be rejected")
+	}
+	if !strings.Contains(err.Error(), "symlink") {
+		t.Fatalf("err = %v, want symlink rejection", err)
+	}
+}
+
+func TestReadRejectsSymlinkInsideWorkingDirToOutside(t *testing.T) {
+	root := t.TempDir()
+	target := filepath.Join(root, "real.txt")
+	if err := os.WriteFile(target, []byte("hello\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	// A symlink to a sibling file still resolves inside the root; Read must
+	// reject it regardless, consistent with Edit's rejectFinalSymlink.
+	link := filepath.Join(root, "alias")
+	if err := os.Symlink(target, link); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := (&Read{}).Execute(context.Background(), agent.ToolContext{WorkingDir: root}, map[string]any{
+		"path": "alias",
+	})
+	if err == nil {
+		t.Fatal("expected symlink to be rejected even when target is inside root")
+	}
+	if !strings.Contains(err.Error(), "symlink") {
+		t.Fatalf("err = %v, want symlink rejection", err)
+	}
+}
@@ -0,0 +1,199 @@
+package tools
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net/url"
+	"strings"
+	"time"
+
+	"github.com/ollama/ollama/agent"
+	"github.com/ollama/ollama/api"
+	internalcloud "github.com/ollama/ollama/internal/cloud"
+)
+
+var (
+	ErrWebSearchAuthRequired = errors.New("web search requires authentication")
+	ErrWebFetchAuthRequired  = errors.New("web fetch requires authentication")
+)
+
+const (
+	maxWebFetchContentRunes = 60_000
+	webSearchTimeout        = 15 * time.Second
+	webFetchTimeout         = 30 * time.Second
+)
+
+type WebSearch struct{}
+
+func (w *WebSearch) Name() string {
+	return "web_search"
+}
+
+func (w *WebSearch) Description() string {
+	return "Search the web for current information that may not be in the model's training data."
+}
+
+func (w *WebSearch) Schema() api.ToolFunction {
+	props := api.NewToolPropertiesMap()
+	props.Set("query", api.ToolProperty{
+		Type:        api.PropertyType{"string"},
+		Description: "The search query to look up on the web.",
+	})
+	return api.ToolFunction{
+		Name:        w.Name(),
+		Description: w.Description(),
+		Parameters: api.ToolFunctionParameters{
+			Type:       "object",
+			Properties: props,
+			Required:   []string{"query"},
+		},
+	}
+}
+
+func (w *WebSearch) RequiresApproval(map[string]any) bool {
+	return true
+}
+
+func (w *WebSearch) Execute(ctx context.Context, _ agent.ToolContext, args map[string]any) (agent.ToolResult, error) {
+	if internalcloud.Disabled() {
+		return agent.ToolResult{}, errors.New(internalcloud.DisabledError("web search is unavailable"))
+	}
+	query, ok := args["query"].(string)
+	if !ok || strings.TrimSpace(query) == "" {
+		return agent.ToolResult{}, fmt.Errorf("query parameter is required")
+	}
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return agent.ToolResult{}, err
+	}
+
+	ctx, cancel := context.WithTimeout(ctx, webSearchTimeout)
+	defer cancel()
+
+	searchResp, err := client.WebSearchExperimental(ctx, &api.WebSearchRequest{Query: query, MaxResults: 5})
+	if err != nil {
+		var authErr api.AuthorizationError
+		if errors.As(err, &authErr) {
+			return agent.ToolResult{}, fmt.Errorf("%w: %s", ErrWebSearchAuthRequired, authErr)
+		}
+		return agent.ToolResult{}, err
+	}
+	if len(searchResp.Results) == 0 {
+		return agent.ToolResult{Content: "No results found for query: " + query}, nil
+	}
+
+	var sb strings.Builder
+	sb.WriteString(fmt.Sprintf("Search results for: %s\n\n", query))
+	for i, result := range searchResp.Results {
+		sb.WriteString(fmt.Sprintf("%d. %s\n", i+1, result.Title))
+		sb.WriteString(fmt.Sprintf("   URL: %s\n", result.URL))
+		if result.Content != "" {
+			content := []rune(result.Content)
+			if len(content) > 300 {
+				content = append(content[:300], []rune("...")...)
+			}
+			sb.WriteString(fmt.Sprintf("   %s\n", string(content)))
+		}
+		sb.WriteByte('\n')
+	}
+	return agent.ToolResult{Content: sb.String()}, nil
+}
+
+type WebFetch struct{}
+
+func (w *WebFetch) Name() string {
+	return "web_fetch"
+}
+
+func (w *WebFetch) Description() string {
+	return "Fetch and extract text content from a web page."
+}
+
+func (w *WebFetch) Schema() api.ToolFunction {
+	props := api.NewToolPropertiesMap()
+	props.Set("url", api.ToolProperty{
+		Type:        api.PropertyType{"string"},
+		Description: "The URL to fetch and extract content from.",
+	})
+	return api.ToolFunction{
+		Name:        w.Name(),
+		Description: w.Description(),
+		Parameters: api.ToolFunctionParameters{
+			Type:       "object",
+			Properties: props,
+			Required:   []string{"url"},
+		},
+	}
+}
+
+func (w *WebFetch) RequiresApproval(map[string]any) bool {
+	return true
+}
+
+func (w *WebFetch) Execute(ctx context.Context, _ agent.ToolContext, args map[string]any) (agent.ToolResult, error) {
+	if internalcloud.Disabled() {
+		return agent.ToolResult{}, errors.New(internalcloud.DisabledError("web fetch is unavailable"))
+	}
+	urlStr, ok := args["url"].(string)
+	if !ok || strings.TrimSpace(urlStr) == "" {
+		return agent.ToolResult{}, fmt.Errorf("url parameter is required")
+	}
+	parsed, err := url.Parse(urlStr)
+	if err != nil {
+		return agent.ToolResult{}, fmt.Errorf("invalid URL: %w", err)
+	}
+	if scheme := strings.ToLower(parsed.Scheme); scheme != "http" && scheme != "https" {
+		return agent.ToolResult{}, fmt.Errorf("unsupported URL scheme %q: only http and https are allowed", parsed.Scheme)
+	}
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return agent.ToolResult{}, err
+	}
+
+	ctx, cancel := context.WithTimeout(ctx, webFetchTimeout)
+	defer cancel()
+
+	fetchResp, err := client.WebFetchExperimental(ctx, &api.WebFetchRequest{URL: urlStr})
+	if err != nil {
+		var authErr api.AuthorizationError
+		if errors.As(err, &authErr) {
+			return agent.ToolResult{}, fmt.Errorf("%w: %s", ErrWebFetchAuthRequired, authErr)
+		}
+		return agent.ToolResult{}, err
+	}
+
+	var sb strings.Builder
+	if fetchResp.Title != "" {
+		sb.WriteString(fmt.Sprintf("Title: %s\n\n", fetchResp.Title))
+	}
+	if fetchResp.Content != "" {
+		sb.WriteString("Content:\n")
+		sb.WriteString(truncateWebFetchContent(fetchResp.Content))
+	} else {
+		sb.WriteString("No content could be extracted from the page.")
+	}
+	return agent.ToolResult{Content: sb.String()}, nil
+}
+
+func truncateWebFetchContent(content string) string {
+	runes := []rune(content)
+	if len(runes) <= maxWebFetchContentRunes {
+		return content
+	}
+	omitted := len(runes) - maxWebFetchContentRunes
+	return string(runes[:maxWebFetchContentRunes]) + fmt.Sprintf(
+		"\n\n[tool output truncated: showing first ~%d tokens; omitted ~%d tokens. Use a narrower request or search query if more detail is needed.]",
+		approximateToolTokensFromRunes(maxWebFetchContentRunes),
+		approximateToolTokensFromRunes(omitted),
+	)
+}
+
+func approximateToolTokensFromRunes(n int) int {
+	if n <= 0 {
+		return 0
+	}
+	return max(1, (n+3)/4)
+}
@@ -0,0 +1,88 @@
+package tools
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	coreagent "github.com/ollama/ollama/agent"
+	"github.com/ollama/ollama/api"
+)
+
+func TestWebToolsRequireApproval(t *testing.T) {
+	if !coreagent.ToolRequiresApproval((&WebSearch{}), map[string]any{"query": "ollama"}) {
+		t.Fatal("web search should require approval")
+	}
+	if !coreagent.ToolRequiresApproval((&WebFetch{}), map[string]any{"url": "https://ollama.com"}) {
+		t.Fatal("web fetch should require approval")
+	}
+}
+
+func TestWebFetchRejectsUnsupportedScheme(t *testing.T) {
+	tests := []struct {
+		name    string
+		url     string
+		wantErr bool
+	}{
+		{name: "file scheme", url: "file:///etc/passwd", wantErr: true},
+		{name: "data scheme", url: "data:text/plain,secret", wantErr: true},
+		{name: "ftp scheme", url: "ftp://example.com/secret", wantErr: true},
+		{name: "http allowed", url: "http://example.com", wantErr: false},
+		{name: "https allowed", url: "https://example.com", wantErr: false},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			_, err := (&WebFetch{}).Execute(t.Context(), coreagent.ToolContext{}, map[string]any{"url": tt.url})
+			if tt.wantErr && err == nil {
+				t.Fatal("expected unsupported scheme to be rejected")
+			}
+			// For allowed schemes we expect an error only from the missing
+			// server/auth path, not from scheme validation. The http/https
+			// cases reach the client and may fail on connection/auth; we only
+			// assert that the error is NOT a scheme error.
+			if !tt.wantErr && err != nil && strings.Contains(err.Error(), "unsupported URL scheme") {
+				t.Fatalf("http/https rejected as unsupported: %v", err)
+			}
+		})
+	}
+}
+
+func TestWebFetchBoundsContentBeforeReturning(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/api/experimental/web_fetch" {
+			t.Fatalf("path = %q, want /api/experimental/web_fetch", r.URL.Path)
+		}
+		var req api.WebFetchRequest
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			t.Fatal(err)
+		}
+		if req.URL != "https://ollama.com" {
+			t.Fatalf("request URL = %q, want https://ollama.com", req.URL)
+		}
+		if err := json.NewEncoder(w).Encode(api.WebFetchResponse{
+			Title:   "Ollama",
+			Content: strings.Repeat("x", maxWebFetchContentRunes+25),
+		}); err != nil {
+			t.Fatal(err)
+		}
+	}))
+	defer ts.Close()
+	t.Setenv("OLLAMA_HOST", ts.URL)
+
+	result, err := (&WebFetch{}).Execute(t.Context(), coreagent.ToolContext{}, map[string]any{
+		"url": "https://ollama.com",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(result.Content, "[tool output truncated: showing first ~") ||
+		!strings.Contains(result.Content, "omitted ~7 tokens") ||
+		!strings.Contains(result.Content, "Use a narrower request or search query") {
+		t.Fatalf("content missing truncation marker: %q", result.Content)
+	}
+	if count := strings.Count(result.Content, "x"); count != maxWebFetchContentRunes {
+		t.Fatalf("captured content count = %d, want %d", count, maxWebFetchContentRunes)
+	}
+}
@@ -78,6 +78,11 @@ type MessagesRequest struct {
 	ToolChoice    *ToolChoice     `json:"tool_choice,omitempty"`
 	Thinking      *ThinkingConfig `json:"thinking,omitempty"`
 	Metadata      *Metadata       `json:"metadata,omitempty"`
+	OutputConfig  *OutputConfig   `json:"output_config,omitempty"`
+}
+
+type OutputConfig struct {
+	Effort string `json:"effort,omitempty"`
 }

 // MessageParam represents a message in the request
@@ -161,7 +166,7 @@ type WebSearchToolResultError struct {

 // ImageSource represents the source of an image
 type ImageSource struct {
-	Type      string `json:"type"` // "base64" or "url"
+	Type      string `json:"type"` // "base64"
 	MediaType string `json:"media_type,omitempty"`
 	Data      string `json:"data,omitempty"`
 	URL       string `json:"url,omitempty"`
@@ -373,9 +378,26 @@ func FromMessagesRequest(r MessagesRequest) (*api.ChatRequest, error) {
 	}

 	var think *api.ThinkValue
+	normalizedEffort := ""
+	if r.OutputConfig != nil {
+		normalizedEffort = strings.ToLower(strings.TrimSpace(r.OutputConfig.Effort))
+		if normalizedEffort == "xhigh" {
+			normalizedEffort = "high"
+		}
+	}
+
 	if r.Thinking != nil && r.Thinking.Type == "enabled" {
 		think = &api.ThinkValue{Value: true}
 	}
+	if r.Thinking != nil && r.Thinking.Type == "disabled" {
+		think = &api.ThinkValue{Value: false}
+	}
+	if think == nil && r.OutputConfig != nil {
+		switch normalizedEffort {
+		case "high", "medium", "low", "max":
+			think = &api.ThinkValue{Value: normalizedEffort}
+		}
+	}

 	stream := r.Stream
 	convertedRequest := &api.ChatRequest{
@@ -425,17 +447,12 @@ func convertMessage(msg MessageParam) ([]api.Message, error) {
 				return nil, errors.New("invalid image source")
 			}

-			if block.Source.Type == "base64" {
-				decoded, err := base64.StdEncoding.DecodeString(block.Source.Data)
-				if err != nil {
-					logutil.Trace("anthropic: invalid base64 image data", "role", role, "error", err)
-					return nil, fmt.Errorf("invalid base64 image data: %w", err)
-				}
-				images = append(images, decoded)
-			} else {
-				logutil.Trace("anthropic: unsupported image source type", "role", role, "source_type", block.Source.Type)
-				return nil, fmt.Errorf("invalid image source type: %s. Only base64 images are supported.", block.Source.Type)
+			decoded, err := resolveImageSource(block.Source)
+			if err != nil {
+				logutil.Trace("anthropic: unsupported image source", "role", role, "source_type", block.Source.Type, "error", err)
+				return nil, err
 			}
+			images = append(images, decoded)

 		case "tool_use":
 			toolUseBlocks++
@@ -457,26 +474,16 @@ func convertMessage(msg MessageParam) ([]api.Message, error) {

 		case "tool_result":
 			toolResultBlocks++
-			var resultContent string
-
-			switch c := block.Content.(type) {
-			case string:
-				resultContent = c
-			case []any:
-				for _, cb := range c {
-					if cbMap, ok := cb.(map[string]any); ok {
-						if cbMap["type"] == "text" {
-							if text, ok := cbMap["text"].(string); ok {
-								resultContent += text
-							}
-						}
-					}
-				}
+			resultContent, resultImages, err := convertToolResultContent(block.Content)
+			if err != nil {
+				logutil.Trace("anthropic: invalid tool_result content", "role", role, "error", err)
+				return nil, err
 			}

 			toolResults = append(toolResults, api.Message{
 				Role:       "tool",
 				Content:    resultContent,
+				Images:     resultImages,
 				ToolCallID: block.ToolUseID,
 			})

@@ -508,6 +515,10 @@ func convertMessage(msg MessageParam) ([]api.Message, error) {
 		}
 	}

+	if role == "user" && len(toolResults) > 0 {
+		messages = append(messages, toolResults...)
+	}
+
 	if textContent.Len() > 0 || len(images) > 0 || len(toolCalls) > 0 || thinking != "" {
 		m := api.Message{
 			Role:      role,
@@ -519,8 +530,10 @@ func convertMessage(msg MessageParam) ([]api.Message, error) {
 		messages = append(messages, m)
 	}

-	// Add tool results as separate messages
-	messages = append(messages, toolResults...)
+	// Add tool results as separate messages.
+	if role != "user" || len(toolResults) == 0 {
+		messages = append(messages, toolResults...)
+	}
 	logutil.Trace("anthropic: converted block message",
 		"role", role,
 		"blocks", len(msg.Content),
@@ -969,6 +982,71 @@ func GenerateMessageID() string {
 	return generateID("msg")
 }

+func resolveImageSource(source *ImageSource) (api.ImageData, error) {
+	if source.Type != "base64" {
+		return nil, fmt.Errorf("invalid image source type: %s. Only base64 images are supported.", source.Type)
+	}
+
+	decoded, err := base64.StdEncoding.DecodeString(source.Data)
+	if err != nil {
+		return nil, fmt.Errorf("invalid base64 image data: %w", err)
+	}
+
+	return decoded, nil
+}
+
+func convertToolResultContent(content any) (string, []api.ImageData, error) {
+	switch c := content.(type) {
+	case nil:
+		return "", nil, nil
+	case string:
+		return c, nil, nil
+	case []any:
+		var text strings.Builder
+		var images []api.ImageData
+
+		for _, cb := range c {
+			cbMap, ok := cb.(map[string]any)
+			if !ok {
+				continue
+			}
+
+			switch cbMap["type"] {
+			case "text":
+				if t, ok := cbMap["text"].(string); ok {
+					text.WriteString(t)
+				}
+			case "image":
+				rawSource, ok := cbMap["source"].(map[string]any)
+				if !ok {
+					return "", nil, errors.New("invalid tool_result image source")
+				}
+
+				var source ImageSource
+				if rawType, ok := rawSource["type"].(string); ok {
+					source.Type = rawType
+				}
+				if rawMediaType, ok := rawSource["media_type"].(string); ok {
+					source.MediaType = rawMediaType
+				}
+				if rawData, ok := rawSource["data"].(string); ok {
+					source.Data = rawData
+				}
+
+				img, err := resolveImageSource(&source)
+				if err != nil {
+					return "", nil, err
+				}
+				images = append(images, img)
+			}
+		}
+
+		return text.String(), images, nil
+	default:
+		return "", nil, nil
+	}
+}
+
 // ptr returns a pointer to the given string value
 func ptr(s string) *string {
 	return &s
@@ -271,6 +271,241 @@ func TestFromMessagesRequest_WithToolResult(t *testing.T) {
 	}
 }

+func TestFromMessagesRequest_WithToolResultImage(t *testing.T) {
+	imgData, _ := base64.StdEncoding.DecodeString(testImage)
+
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "user",
+				Content: []ContentBlock{
+					{
+						Type:      "tool_result",
+						ToolUseID: "call_img",
+						Content: []any{
+							map[string]any{"type": "text", "text": "Attached image"},
+							map[string]any{
+								"type": "image",
+								"source": map[string]any{
+									"type":       "base64",
+									"media_type": "image/png",
+									"data":       testImage,
+								},
+							},
+						},
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 1 {
+		t.Fatalf("expected 1 message, got %d", len(result.Messages))
+	}
+
+	msg := result.Messages[0]
+	if msg.Role != "tool" {
+		t.Errorf("expected role 'tool', got %q", msg.Role)
+	}
+	if msg.ToolCallID != "call_img" {
+		t.Errorf("expected tool_call_id 'call_img', got %q", msg.ToolCallID)
+	}
+	if msg.Content != "Attached image" {
+		t.Errorf("unexpected content: %q", msg.Content)
+	}
+	if len(msg.Images) != 1 {
+		t.Fatalf("expected 1 image, got %d", len(msg.Images))
+	}
+	if string(msg.Images[0]) != string(imgData) {
+		t.Error("image data mismatch")
+	}
+}
+
+func TestFromMessagesRequest_WithToolResultFollowedByUserText(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "assistant",
+				Content: []ContentBlock{
+					{
+						Type:  "tool_use",
+						ID:    "call_read",
+						Name:  "Read",
+						Input: makeArgs("file_path", "/Users/hoyyeva/Desktop/aaa.png"),
+					},
+				},
+			},
+			{
+				Role: "user",
+				Content: []ContentBlock{
+					{
+						Type:      "tool_result",
+						ToolUseID: "call_read",
+						Content:   "Read image (311.5KB)",
+					},
+					{
+						Type: "text",
+						Text: ptr("Please describe it."),
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 3 {
+		t.Fatalf("expected 3 messages, got %d", len(result.Messages))
+	}
+
+	if result.Messages[1].Role != "tool" {
+		t.Fatalf("expected second message to be tool, got %q", result.Messages[1].Role)
+	}
+	if result.Messages[1].ToolCallID != "call_read" {
+		t.Fatalf("expected tool_call_id 'call_read', got %q", result.Messages[1].ToolCallID)
+	}
+	if result.Messages[2].Role != "user" {
+		t.Fatalf("expected third message to be user, got %q", result.Messages[2].Role)
+	}
+	if result.Messages[2].Content != "Please describe it." {
+		t.Fatalf("unexpected user content: %q", result.Messages[2].Content)
+	}
+}
+
+func TestFromMessagesRequest_WithOutputConfigEffort(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "gemma4",
+		MaxTokens: 32000,
+		Messages: []MessageParam{
+			{
+				Role:    "user",
+				Content: textContent("Describe the image."),
+			},
+		},
+		OutputConfig: &OutputConfig{
+			Effort: "high",
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Think == nil {
+		t.Fatal("expected think to be set from output_config.effort")
+	}
+
+	if got := result.Think.String(); got != "high" {
+		t.Fatalf("expected think level 'high', got %q", got)
+	}
+}
+
+func TestFromMessagesRequest_WithOutputConfigEffortXHighMapsToHigh(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "gemma4",
+		MaxTokens: 32000,
+		Messages: []MessageParam{
+			{
+				Role:    "user",
+				Content: textContent("Describe the image."),
+			},
+		},
+		OutputConfig: &OutputConfig{
+			Effort: "xhigh",
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Think == nil {
+		t.Fatal("expected think to be set from output_config.effort")
+	}
+
+	if got := result.Think.String(); got != "high" {
+		t.Fatalf("expected think level 'high' for xhigh effort, got %q", got)
+	}
+}
+
+func TestFromMessagesRequest_ThinkingDisabledOverridesOutputConfigEffort(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "gemma4",
+		MaxTokens: 32000,
+		Messages: []MessageParam{
+			{
+				Role:    "user",
+				Content: textContent("Describe the image."),
+			},
+		},
+		Thinking: &ThinkingConfig{
+			Type: "disabled",
+		},
+		OutputConfig: &OutputConfig{
+			Effort: "high",
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Think == nil {
+		t.Fatal("expected think to be set")
+	}
+
+	if got := result.Think.Value; got != false {
+		t.Fatalf("expected think=false when thinking is disabled, got %v", got)
+	}
+}
+
+func TestFromMessagesRequest_ThinkingAdaptiveUsesOutputConfigEffort(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "gemma4",
+		MaxTokens: 32000,
+		Messages: []MessageParam{
+			{
+				Role:    "user",
+				Content: textContent("Describe the image."),
+			},
+		},
+		Thinking: &ThinkingConfig{
+			Type: "adaptive",
+		},
+		OutputConfig: &OutputConfig{
+			Effort: "high",
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Think == nil {
+		t.Fatal("expected think to be set from output_config.effort")
+	}
+
+	if got := result.Think.String(); got != "high" {
+		t.Fatalf("expected think level 'high' for adaptive thinking, got %q", got)
+	}
+}
+
 func TestFromMessagesRequest_WithTools(t *testing.T) {
 	req := MessagesRequest{
 		Model:     "test-model",
@@ -259,6 +259,10 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		}
 	}

+	if err := scanner.Err(); err != nil {
+		return err
+	}
+
 	return nil
 }

@@ -368,6 +372,16 @@ func (c *Client) List(ctx context.Context) (*ListResponse, error) {
 	return &lr, nil
 }

+// ModelRecommendationsExperimental lists model recommendations from the local
+// server's experimental recommendations endpoint.
+func (c *Client) ModelRecommendationsExperimental(ctx context.Context) (*ModelRecommendationsResponse, error) {
+	var resp ModelRecommendationsResponse
+	if err := c.do(ctx, http.MethodGet, "/api/experimental/model-recommendations", nil, &resp); err != nil {
+		return nil, err
+	}
+	return &resp, nil
+}
+
 // ListRunning lists running models.
 func (c *Client) ListRunning(ctx context.Context) (*ProcessResponse, error) {
 	var lr ProcessResponse
@@ -459,6 +473,26 @@ func (c *Client) CloudStatusExperimental(ctx context.Context) (*StatusResponse,
 	return &status, nil
 }

+// WebSearchExperimental searches the web through the local server's
+// experimental web search endpoint.
+func (c *Client) WebSearchExperimental(ctx context.Context, req *WebSearchRequest) (*WebSearchResponse, error) {
+	var resp WebSearchResponse
+	if err := c.do(ctx, http.MethodPost, "/api/experimental/web_search", req, &resp); err != nil {
+		return nil, err
+	}
+	return &resp, nil
+}
+
+// WebFetchExperimental fetches web page content through the local server's
+// experimental web fetch endpoint.
+func (c *Client) WebFetchExperimental(ctx context.Context, req *WebFetchRequest) (*WebFetchResponse, error) {
+	var resp WebFetchResponse
+	if err := c.do(ctx, http.MethodPost, "/api/experimental/web_fetch", req, &resp); err != nil {
+		return nil, err
+	}
+	return &resp, nil
+}
+
 // Signout will signout a client for a local ollama server.
 func (c *Client) Signout(ctx context.Context) error {
 	return c.do(ctx, http.MethodPost, "/api/signout", nil, nil)
@@ -3,6 +3,7 @@ package api
 import (
 	"encoding/json"
 	"fmt"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"net/url"
@@ -192,6 +193,35 @@ func TestClientStream(t *testing.T) {
 	}
 }

+func TestClientStreamReportsReadErrors(t *testing.T) {
+	client := NewClient(
+		&url.URL{Scheme: "http", Host: "example.com"},
+		&http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+			body := failingReader{
+				data: []byte(`{"message":{"content":"partial"}}` + "\n"),
+				err:  io.ErrUnexpectedEOF,
+			}
+
+			return &http.Response{
+				StatusCode: http.StatusOK,
+				Status:     "200 OK",
+				Body:       io.NopCloser(&body),
+				Header:     make(http.Header),
+			}, nil
+		})},
+	)
+
+	err := client.stream(t.Context(), http.MethodPost, "/api/chat", nil, func([]byte) error {
+		return nil
+	})
+	if err == nil {
+		t.Fatal("expected stream read error")
+	}
+	if !strings.Contains(err.Error(), io.ErrUnexpectedEOF.Error()) {
+		t.Fatalf("expected unexpected EOF, got %v", err)
+	}
+}
+
 func TestClientDo(t *testing.T) {
 	testCases := []struct {
 		name           string
@@ -320,3 +350,99 @@ func TestClientDo(t *testing.T) {
 		})
 	}
 }
+
+func TestClientWebSearchExperimentalUsesLocalRoute(t *testing.T) {
+	var gotPath string
+	var gotMethod string
+	var gotRequest WebSearchRequest
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		gotMethod = r.Method
+		if err := json.NewDecoder(r.Body).Decode(&gotRequest); err != nil {
+			t.Fatal(err)
+		}
+		if err := json.NewEncoder(w).Encode(WebSearchResponse{
+			Results: []WebSearchResult{{Title: "Ollama", URL: "https://ollama.com", Content: "models"}},
+		}); err != nil {
+			t.Fatal(err)
+		}
+	}))
+	defer ts.Close()
+
+	client := NewClient(&url.URL{Scheme: "http", Host: ts.Listener.Addr().String()}, http.DefaultClient)
+	resp, err := client.WebSearchExperimental(t.Context(), &WebSearchRequest{Query: "ollama", MaxResults: 3})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if gotMethod != http.MethodPost {
+		t.Fatalf("method = %q, want POST", gotMethod)
+	}
+	if gotPath != "/api/experimental/web_search" {
+		t.Fatalf("path = %q, want /api/experimental/web_search", gotPath)
+	}
+	if gotRequest.Query != "ollama" || gotRequest.MaxResults != 3 {
+		t.Fatalf("request = %#v", gotRequest)
+	}
+	if len(resp.Results) != 1 || resp.Results[0].Title != "Ollama" {
+		t.Fatalf("response = %#v", resp)
+	}
+}
+
+func TestClientWebFetchExperimentalUsesLocalRoute(t *testing.T) {
+	var gotPath string
+	var gotMethod string
+	var gotRequest WebFetchRequest
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		gotMethod = r.Method
+		if err := json.NewDecoder(r.Body).Decode(&gotRequest); err != nil {
+			t.Fatal(err)
+		}
+		if err := json.NewEncoder(w).Encode(WebFetchResponse{
+			Title:   "Ollama",
+			Content: "models",
+			Links:   []string{"https://ollama.com/library"},
+		}); err != nil {
+			t.Fatal(err)
+		}
+	}))
+	defer ts.Close()
+
+	client := NewClient(&url.URL{Scheme: "http", Host: ts.Listener.Addr().String()}, http.DefaultClient)
+	resp, err := client.WebFetchExperimental(t.Context(), &WebFetchRequest{URL: "https://ollama.com"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if gotMethod != http.MethodPost {
+		t.Fatalf("method = %q, want POST", gotMethod)
+	}
+	if gotPath != "/api/experimental/web_fetch" {
+		t.Fatalf("path = %q, want /api/experimental/web_fetch", gotPath)
+	}
+	if gotRequest.URL != "https://ollama.com" {
+		t.Fatalf("request = %#v", gotRequest)
+	}
+	if resp.Title != "Ollama" || resp.Content != "models" {
+		t.Fatalf("response = %#v", resp)
+	}
+}
+
+type roundTripFunc func(*http.Request) (*http.Response, error)
+
+func (f roundTripFunc) RoundTrip(req *http.Request) (*http.Response, error) {
+	return f(req)
+}
+
+type failingReader struct {
+	data []byte
+	err  error
+}
+
+func (r *failingReader) Read(p []byte) (int, error) {
+	if len(r.data) > 0 {
+		n := copy(p, r.data)
+		r.data = r.data[n:]
+		return n, nil
+	}
+	return 0, r.err
+}
@@ -436,6 +436,7 @@ type ToolProperty struct {
 	Description string             `json:"description,omitempty"`
 	Enum        []any              `json:"enum,omitempty"`
 	Properties  *ToolPropertiesMap `json:"properties,omitempty"`
+	Required    []string           `json:"required,omitempty"`
 }

 // ToTypeScriptType converts a ToolProperty to a TypeScript type string
@@ -599,12 +600,13 @@ type Options struct {

 // Runner options which must be set when the model is loaded into memory
 type Runner struct {
-	NumCtx    int   `json:"num_ctx,omitempty"`
-	NumBatch  int   `json:"num_batch,omitempty"`
-	NumGPU    int   `json:"num_gpu,omitempty"`
-	MainGPU   int   `json:"main_gpu,omitempty"`
-	UseMMap   *bool `json:"use_mmap,omitempty"`
-	NumThread int   `json:"num_thread,omitempty"`
+	NumCtx          int   `json:"num_ctx,omitempty"`
+	NumBatch        int   `json:"num_batch,omitempty"`
+	NumGPU          int   `json:"num_gpu,omitempty"`
+	MainGPU         *int  `json:"main_gpu,omitempty"`
+	UseMMap         *bool `json:"use_mmap,omitempty"`
+	NumThread       int   `json:"num_thread,omitempty"`
+	DraftNumPredict int   `json:"draft_num_predict,omitempty"`
 }

 // EmbedRequest is the request passed to [Client.Embed].
@@ -671,6 +673,9 @@ type CreateRequest struct {
 	// Quantize is the quantization format for the model; leave blank to not change the quantization level.
 	Quantize string `json:"quantize,omitempty"`

+	// DraftQuantize is the quantization format for the draft model.
+	DraftQuantize string `json:"draft_quantize,omitempty"`
+
 	// From is the name of the model or file to use as the source.
 	From string `json:"from,omitempty"`

@@ -680,6 +685,9 @@ type CreateRequest struct {
 	// Files is a map of files include when creating the model.
 	Files map[string]string `json:"files,omitempty"`

+	// DraftFiles is a map of draft model files to include when creating the model.
+	DraftFiles map[string]string `json:"draft_files,omitempty"`
+
 	// Adapters is a map of LoRA adapters to include when creating the model.
 	Adapters map[string]string `json:"adapters,omitempty"`

@@ -801,6 +809,21 @@ type ListResponse struct {
 	Models []ListModelResponse `json:"models"`
 }

+// ModelRecommendationsResponse is the response from [Client.ModelRecommendationsExperimental].
+type ModelRecommendationsResponse struct {
+	Recommendations []ModelRecommendation `json:"recommendations"`
+}
+
+// ModelRecommendation is a single recommendation entry in [ModelRecommendationsResponse].
+type ModelRecommendation struct {
+	Model           string `json:"model"`
+	Description     string `json:"description"`
+	ContextLength   int    `json:"context_length,omitempty"`
+	MaxOutputTokens int    `json:"max_output_tokens,omitempty"`
+	VRAMBytes       int64  `json:"vram_bytes,omitempty"`
+	RequiredPlan    string `json:"required_plan,omitempty"`
+}
+
 // ProcessResponse is the response from [Client.Process].
 type ProcessResponse struct {
 	Models []ProcessModelResponse `json:"models"`
@@ -808,14 +831,15 @@ type ProcessResponse struct {

 // ListModelResponse is a single model description in [ListResponse].
 type ListModelResponse struct {
-	Name        string       `json:"name"`
-	Model       string       `json:"model"`
-	RemoteModel string       `json:"remote_model,omitempty"`
-	RemoteHost  string       `json:"remote_host,omitempty"`
-	ModifiedAt  time.Time    `json:"modified_at"`
-	Size        int64        `json:"size"`
-	Digest      string       `json:"digest"`
-	Details     ModelDetails `json:"details,omitempty"`
+	Name         string             `json:"name"`
+	Model        string             `json:"model"`
+	RemoteModel  string             `json:"remote_model,omitempty"`
+	RemoteHost   string             `json:"remote_host,omitempty"`
+	ModifiedAt   time.Time          `json:"modified_at"`
+	Size         int64              `json:"size"`
+	Digest       string             `json:"digest"`
+	Details      ModelDetails       `json:"details,omitempty"`
+	Capabilities []model.Capability `json:"capabilities,omitempty"`
 }

 // ProcessModelResponse is a single model description in [ProcessResponse].
@@ -844,6 +868,36 @@ type StatusResponse struct {
 	Cloud CloudStatus `json:"cloud"`
 }

+// WebSearchRequest is the request for [Client.WebSearchExperimental].
+type WebSearchRequest struct {
+	Query      string `json:"query"`
+	MaxResults int    `json:"max_results,omitempty"`
+}
+
+// WebSearchResult is a single result from [Client.WebSearchExperimental].
+type WebSearchResult struct {
+	Title   string `json:"title"`
+	URL     string `json:"url"`
+	Content string `json:"content"`
+}
+
+// WebSearchResponse is the response from [Client.WebSearchExperimental].
+type WebSearchResponse struct {
+	Results []WebSearchResult `json:"results"`
+}
+
+// WebFetchRequest is the request for [Client.WebFetchExperimental].
+type WebFetchRequest struct {
+	URL string `json:"url"`
+}
+
+// WebFetchResponse is the response from [Client.WebFetchExperimental].
+type WebFetchResponse struct {
+	Title   string   `json:"title"`
+	Content string   `json:"content"`
+	Links   []string `json:"links,omitempty"`
+}
+
 // GenerateResponse is the response passed into [GenerateResponseFunc].
 type GenerateResponse struct {
 	// Model is the model name that generated the response.
@@ -908,6 +962,8 @@ type ModelDetails struct {
 	Families          []string `json:"families"`
 	ParameterSize     string   `json:"parameter_size"`
 	QuantizationLevel string   `json:"quantization_level"`
+	ContextLength     int      `json:"context_length,omitempty"`
+	EmbeddingLength   int      `json:"embedding_length,omitempty"`
 }

 // UserResponse provides information about a user.
@@ -1030,14 +1086,25 @@ func (opts *Options) FromMap(m map[string]any) error {
 				}
 				field.Set(reflect.ValueOf(slice))
 			case reflect.Pointer:
-				var b bool
-				if field.Type() == reflect.TypeOf(&b) {
+				switch field.Type().Elem().Kind() {
+				case reflect.Bool:
 					val, ok := val.(bool)
 					if !ok {
 						return fmt.Errorf("option %q must be of type boolean", key)
 					}
 					field.Set(reflect.ValueOf(&val))
-				} else {
+				case reflect.Int:
+					var i int
+					switch t := val.(type) {
+					case int64:
+						i = int(t)
+					case float64:
+						i = int(t)
+					default:
+						return fmt.Errorf("option %q must be of type integer", key)
+					}
+					field.Set(reflect.ValueOf(&i))
+				default:
 					return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type())
 				}
 			default:
@@ -1070,16 +1137,17 @@ func DefaultOptions() Options {

 		Runner: Runner{
 			// options set when the model is loaded
-			NumCtx:    int(envconfig.ContextLength()),
-			NumBatch:  512,
-			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
-			NumThread: 0,  // let the runtime decide
-			UseMMap:   nil,
+			NumCtx:          int(envconfig.ContextLength()),
+			NumBatch:        512,
+			NumGPU:          -1, // -1 here indicates that NumGPU should be set dynamically
+			NumThread:       0,  // let the runtime decide
+			DraftNumPredict: 4,
+			UseMMap:         nil,
 		},
 	}
 }

-// ThinkValue represents a value that can be a boolean or a string ("high", "medium", "low")
+// ThinkValue represents a value that can be a boolean or a string ("high", "medium", "low", "max")
 type ThinkValue struct {
 	// Value can be a bool or string
 	Value interface{}
@@ -1095,7 +1163,7 @@ func (t *ThinkValue) IsValid() bool {
 	case bool:
 		return true
 	case string:
-		return v == "high" || v == "medium" || v == "low"
+		return v == "high" || v == "medium" || v == "low" || v == "max"
 	default:
 		return false
 	}
@@ -1129,8 +1197,8 @@ func (t *ThinkValue) Bool() bool {
 	case bool:
 		return v
 	case string:
-		// Any string value ("high", "medium", "low") means thinking is enabled
-		return v == "high" || v == "medium" || v == "low"
+		// Any string value ("high", "medium", "low", "max") means thinking is enabled
+		return v == "high" || v == "medium" || v == "low" || v == "max"
 	default:
 		return false
 	}
@@ -1168,14 +1236,14 @@ func (t *ThinkValue) UnmarshalJSON(data []byte) error {
 	var s string
 	if err := json.Unmarshal(data, &s); err == nil {
 		// Validate string values
-		if s != "high" && s != "medium" && s != "low" {
-			return fmt.Errorf("invalid think value: %q (must be \"high\", \"medium\", \"low\", true, or false)", s)
+		if s != "high" && s != "medium" && s != "low" && s != "max" {
+			return fmt.Errorf("invalid think value: %q (must be \"high\", \"medium\", \"low\", \"max\", true, or false)", s)
 		}
 		t.Value = s
 		return nil
 	}

-	return fmt.Errorf("think must be a boolean or string (\"high\", \"medium\", \"low\", true, or false)")
+	return fmt.Errorf("think must be a boolean or string (\"high\", \"medium\", \"low\", \"max\", true, or false)")
 }

 // MarshalJSON implements json.Marshaler
@@ -1278,14 +1346,20 @@ func FormatParams(params map[string][]string) (map[string]any, error) {
 					// TODO: only string slices are supported right now
 					out[key] = vals
 				case reflect.Pointer:
-					var b bool
-					if field.Type() == reflect.TypeOf(&b) {
+					switch field.Type().Elem().Kind() {
+					case reflect.Bool:
 						boolVal, err := strconv.ParseBool(vals[0])
 						if err != nil {
 							return nil, fmt.Errorf("invalid bool value %s", vals)
 						}
 						out[key] = &boolVal
-					} else {
+					case reflect.Int:
+						intVal, err := strconv.ParseInt(vals[0], 10, 64)
+						if err != nil {
+							return nil, fmt.Errorf("invalid int value %s", vals)
+						}
+						out[key] = intVal
+					default:
 						return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
 					}
 				default:
@@ -20,6 +20,10 @@ func testPropsMap(m map[string]ToolProperty) *ToolPropertiesMap {
 	return props
 }

+func testIntPtr(v int) *int {
+	return &v
+}
+
 // testArgs creates ToolCallFunctionArguments from a map (convenience function for tests, order not preserved)
 func testArgs(m map[string]any) ToolCallFunctionArguments {
 	args := NewToolCallFunctionArguments()
@@ -168,6 +172,47 @@ func TestUseMmapParsingFromJSON(t *testing.T) {
 	}
 }

+func TestMainGPUParsingFromJSON(t *testing.T) {
+	tests := []struct {
+		name    string
+		req     string
+		wantGPU *int
+	}{
+		{
+			name: "Undefined",
+			req:  `{}`,
+		},
+		{
+			name:    "Zero",
+			req:     `{ "main_gpu": 0 }`,
+			wantGPU: testIntPtr(0),
+		},
+		{
+			name:    "Nonzero",
+			req:     `{ "main_gpu": 1 }`,
+			wantGPU: testIntPtr(1),
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var oMap map[string]any
+			err := json.Unmarshal([]byte(test.req), &oMap)
+			require.NoError(t, err)
+
+			opts := DefaultOptions()
+			err = opts.FromMap(oMap)
+			require.NoError(t, err)
+
+			if test.wantGPU == nil {
+				assert.Nil(t, opts.MainGPU)
+			} else if assert.NotNil(t, opts.MainGPU) {
+				assert.Equal(t, *test.wantGPU, *opts.MainGPU)
+			}
+		})
+	}
+}
+
 func TestUseMmapFormatParams(t *testing.T) {
 	tr := true
 	fa := false
@@ -232,6 +277,12 @@ func TestUseMmapFormatParams(t *testing.T) {
 	}
 }

+func TestMainGPUFormatParams(t *testing.T) {
+	resp, err := FormatParams(map[string][]string{"main_gpu": {"0"}})
+	require.NoError(t, err)
+	assert.Equal(t, int64(0), resp["main_gpu"])
+}
+
 func TestMessage_UnmarshalJSON(t *testing.T) {
 	tests := []struct {
 		input    string
@@ -495,6 +546,11 @@ func TestThinking_UnmarshalJSON(t *testing.T) {
 			input:            `{ "think": "low" }`,
 			expectedThinking: &ThinkValue{Value: "low"},
 		},
+		{
+			name:             "string_max",
+			input:            `{ "think": "max" }`,
+			expectedThinking: &ThinkValue{Value: "max"},
+		},
 		{
 			name:             "invalid_string",
 			input:            `{ "think": "invalid" }`,
@@ -157,10 +157,6 @@ func main() {
 		}
 	}

-	if u := os.Getenv("OLLAMA_UPDATE_URL"); u != "" {
-		updater.UpdateCheckURLBase = u
-	}
-
 	// Detect if this is a first start after an upgrade, in
 	// which case we need to do some cleanup
 	var skipMove bool
@@ -14,6 +14,7 @@
 #define MyAppPublisher "Ollama"
 #define MyAppURL "https://ollama.com/"
 #define MyAppExeName "ollama app.exe"
+#define LlamaServerExeName "llama-server.exe"
 #define MyIcon ".\assets\app.ico"

 [Setup]
@@ -90,9 +91,8 @@ DialogFontSize=12
 [Files]
 #if FileExists("..\dist\windows-ollama-app-amd64.exe")
 Source: "..\dist\windows-ollama-app-amd64.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ;Check: not IsArm64();  Flags: ignoreversion 64bit; BeforeInstall: TaskKill('{#MyAppExeName}')
-Source: "..\dist\windows-amd64\vc_redist.x64.exe"; DestDir: "{tmp}"; Check: not IsArm64() and vc_redist_needed(); Flags: deleteafterinstall
 Source: "..\dist\windows-amd64\ollama.exe"; DestDir: "{app}"; Check: not IsArm64(); Flags: ignoreversion 64bit; BeforeInstall: TaskKill('ollama.exe')
-Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: not IsArm64(); Flags: ignoreversion 64bit recursesubdirs
+Source: "..\dist\windows-amd64\lib\ollama\*"; Excludes: "\mlx_*\*"; DestDir: "{app}\lib\ollama\"; Check: not IsArm64(); Flags: ignoreversion 64bit recursesubdirs
 #endif

 ; For local development, rely on binary compatibility at runtime since we can't cross compile
@@ -103,9 +103,11 @@ Source: "..\dist\windows-ollama-app-amd64.exe"; DestDir: "{app}"; DestName: "{#M
 #endif

 #if FileExists("..\dist\windows-arm64\ollama.exe")
-Source: "..\dist\windows-arm64\vc_redist.arm64.exe"; DestDir: "{tmp}"; Check: IsArm64() and vc_redist_needed(); Flags: deleteafterinstall
 Source: "..\dist\windows-arm64\ollama.exe"; DestDir: "{app}"; Check: IsArm64(); Flags: ignoreversion 64bit; BeforeInstall: TaskKill('ollama.exe')
 #endif
+#if DirExists("..\dist\windows-arm64\lib\ollama")
+Source: "..\dist\windows-arm64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: IsArm64(); Flags: ignoreversion 64bit recursesubdirs
+#endif

 Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion

@@ -118,12 +120,6 @@ Name: "{userprograms}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFile
 Type: files; Name: "{%LOCALAPPDATA}\Ollama\updates"

 [Run]
-#if DirExists("..\dist\windows-arm64")
-Filename: "{tmp}\vc_redist.arm64.exe"; Parameters: "/install /passive /norestart"; Check: IsArm64() and vc_redist_needed(); StatusMsg: "Installing VC++ Redistributables..."; Flags: waituntilterminated
-#endif
-#if DirExists("..\dist\windows-amd64")
-Filename: "{tmp}\vc_redist.x64.exe"; Parameters: "/install /passive /norestart"; Check: not IsArm64() and vc_redist_needed(); StatusMsg: "Installing VC++ Redistributables..."; Flags: waituntilterminated
-#endif
 Filename: "{cmd}"; Parameters: "/C set PATH={app};%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden

 [UninstallRun]
@@ -131,6 +127,7 @@ Filename: "{cmd}"; Parameters: "/C set PATH={app};%PATH% & ""{app}\{#MyAppExeNam
 ; Filename: "{cmd}"; Parameters: "/C ""taskkill /im ollama.exe /f /t"; Flags: runhidden
 Filename: "taskkill"; Parameters: "/im ""{#MyAppExeName}"" /f /t"; Flags: runhidden
 Filename: "taskkill"; Parameters: "/im ""ollama.exe"" /f /t"; Flags: runhidden
+Filename: "taskkill"; Parameters: "/im ""{#LlamaServerExeName}"" /f /t"; Flags: runhidden
 ; HACK!  need to give the server and app enough time to exit
 ; TODO - convert this to a Pascal code script so it waits until they're no longer running, then completes
 Filename: "{cmd}"; Parameters: "/c timeout 5"; Flags: runhidden
@@ -184,46 +181,6 @@ begin
  Result := Pos(';' + ExpandConstant(Param) + ';', ';' + OrigPath + ';') = 0;
 end;

-{ --- VC Runtime libraries discovery code - Only install vc_redist if it isn't already installed ----- }
-const VCRTL_MIN_V1 = 14;
-const VCRTL_MIN_V2 = 40;
-const VCRTL_MIN_V3 = 33807;
-const VCRTL_MIN_V4 = 0;
-
- // check if the minimum required vc redist is installed (by looking the registry)
-function vc_redist_needed (): Boolean;
-var
-  sRegKey: string;
-  v1: Cardinal;
-  v2: Cardinal;
-  v3: Cardinal;
-  v4: Cardinal;
-begin
-  if (IsArm64()) then begin
-    sRegKey := 'SOFTWARE\WOW6432Node\Microsoft\VisualStudio\14.0\VC\Runtimes\arm64';
-  end else begin
-    sRegKey := 'SOFTWARE\Microsoft\VisualStudio\14.0\VC\Runtimes\x64';
-  end;
-  if (RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Major', v1)  and
-      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Minor', v2) and
-      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Bld', v3) and
-      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'RBld', v4)) then
-  begin
-    Log ('VC Redist version: ' + IntToStr (v1) +
-        '.' + IntToStr (v2) + '.' + IntToStr (v3) +
-        '.' + IntToStr (v4));
-    { Version info was found. Return true if later or equal to our
-       minimal required version RTL_MIN_Vx }
-    Result := not (
-        (v1 > VCRTL_MIN_V1) or ((v1 = VCRTL_MIN_V1) and
-         ((v2 > VCRTL_MIN_V2) or ((v2 = VCRTL_MIN_V2) and
-          ((v3 > VCRTL_MIN_V3) or ((v3 = VCRTL_MIN_V3) and
-           (v4 >= VCRTL_MIN_V4)))))));
-  end
-  else
-    Result := TRUE;
-end;
-
 function GetDirSize(Path: String): Int64;
 var
  FindRec: TFindRec;
@@ -370,5 +327,8 @@ procedure TaskKill(FileName: String);
 var
  ResultCode: Integer;
 begin
-    Exec('taskkill.exe', '/f /im ' + '"' + FileName + '"', '', SW_HIDE, ewWaitUntilTerminated, ResultCode);
+    Exec('taskkill.exe', '/f /t /im ' + '"' + FileName + '"', '', SW_HIDE, ewWaitUntilTerminated, ResultCode);
+    if FileName <> '{#LlamaServerExeName}' then begin
+      Exec('taskkill.exe', '/f /t /im "{#LlamaServerExeName}"', '', SW_HIDE, ewWaitUntilTerminated, ResultCode);
+    end;
 end;
@@ -83,6 +83,29 @@ func resolvePath(name string) string {
 	return name
 }

+func ollamaServeArgs(args []string) bool {
+	if len(args) < 2 {
+		return false
+	}
+
+	switch strings.Trim(filepath.Base(args[0]), `"`) {
+	case "ollama", "ollama.exe":
+	default:
+		return false
+	}
+
+	for _, rawArg := range args[1:] {
+		arg := strings.Trim(rawArg, `"`)
+		if strings.HasPrefix(arg, "-") {
+			continue
+		}
+
+		return arg == "serve" || arg == "start"
+	}
+
+	return false
+}
+
 // cleanup checks the pid file for a running ollama process
 // and shuts it down gracefully if it is running
 func cleanup() error {
@@ -205,6 +205,63 @@ func TestServerCmdCloudSettingEnv(t *testing.T) {
 	}
 }

+func TestOllamaServeArgs(t *testing.T) {
+	tests := []struct {
+		name string
+		args []string
+		want bool
+	}{
+		{
+			name: "system ollama serve",
+			args: []string{"ollama", "serve"},
+			want: true,
+		},
+		{
+			name: "relative path ollama serve",
+			args: []string{"./ollama", "serve"},
+			want: true,
+		},
+		{
+			name: "serve after other flags",
+			args: []string{"./ollama", "--verbose", "serve"},
+			want: true,
+		},
+		{
+			name: "start alias",
+			args: []string{"ollama", "start"},
+			want: true,
+		},
+		{
+			name: "launch command",
+			args: []string{"ollama", "launch", "opencode"},
+			want: false,
+		},
+		{
+			name: "run command with model named serve",
+			args: []string{"ollama", "run", "serve"},
+			want: false,
+		},
+		{
+			name: "launch command with serve in passthrough args",
+			args: []string{"ollama", "launch", "codex", "--", "-p", "serve"},
+			want: false,
+		},
+		{
+			name: "different executable",
+			args: []string{"go", "run", "serve"},
+			want: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := ollamaServeArgs(tt.args); got != tt.want {
+				t.Fatalf("ollamaServeArgs(%v) = %v, want %v", tt.args, got, tt.want)
+			}
+		})
+	}
+}
+
 func TestGetInferenceInfo(t *testing.T) {
 	tests := []struct {
 		name             string
@@ -46,7 +46,17 @@ func terminated(pid int) (bool, error) {
 	return false, nil
 }

-// reapServers kills all ollama processes except our own
+func ollamaServeProcess(pid int) bool {
+	output, err := exec.Command("ps", "-p", strconv.Itoa(pid), "-o", "args=").Output()
+	if err != nil {
+		slog.Debug("failed to inspect ollama process", "pid", pid, "err", err)
+		return false
+	}
+
+	return ollamaServeArgs(strings.Fields(strings.TrimSpace(string(output))))
+}
+
+// reapServers kills external ollama serve processes except our own.
 func reapServers() error {
 	// Get our own PID to avoid killing ourselves
 	currentPID := os.Getpid()
@@ -82,6 +92,9 @@ func reapServers() error {
 		if pid == currentPID {
 			continue
 		}
+		if !ollamaServeProcess(pid) {
+			continue
+		}

 		proc, err := os.FindProcess(pid)
 		if err != nil {
@@ -101,7 +101,29 @@ func terminated(pid int) (bool, error) {
 	return true, nil
 }

-// reapServers kills all ollama processes except our own
+func ollamaServeProcess(pid int) bool {
+	cmd := exec.Command("wmic", "process", "where", fmt.Sprintf("ProcessId=%d", pid), "get", "CommandLine", "/value")
+	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: true}
+	output, err := cmd.Output()
+	if err != nil {
+		slog.Debug("failed to inspect ollama process", "pid", pid, "err", err)
+		return false
+	}
+
+	for _, line := range strings.Split(string(output), "\n") {
+		line = strings.TrimSpace(line)
+		commandLine, ok := strings.CutPrefix(line, "CommandLine=")
+		if !ok {
+			continue
+		}
+
+		return ollamaServeArgs(strings.Fields(strings.ToLower(commandLine)))
+	}
+
+	return false
+}
+
+// reapServers kills external ollama serve processes except our own.
 func reapServers() error {
 	// Get current process ID to avoid killing ourselves
 	currentPID := os.Getpid()
@@ -138,8 +160,11 @@ func reapServers() error {
 		if pid == currentPID {
 			continue
 		}
+		if !ollamaServeProcess(pid) {
+			continue
+		}

-		cmd := exec.Command("taskkill", "/F", "/PID", pidStr)
+		cmd := exec.Command("taskkill", "/F", "/T", "/PID", pidStr)
 		if err := cmd.Run(); err != nil {
 			slog.Warn("failed to kill ollama process", "pid", pid, "err", err)
 		}
@@ -14,7 +14,7 @@ import (

 // currentSchemaVersion defines the current database schema version.
 // Increment this when making schema changes that require migrations.
-const currentSchemaVersion = 15
+const currentSchemaVersion = 16

 // database wraps the SQLite connection.
 // SQLite handles its own locking for concurrent access:
@@ -82,6 +82,7 @@ func (db *database) init() error {
 		websearch_enabled BOOLEAN NOT NULL DEFAULT 0,
 		selected_model TEXT NOT NULL DEFAULT '',
 		sidebar_open BOOLEAN NOT NULL DEFAULT 0,
+		last_home_view TEXT NOT NULL DEFAULT 'launch',
 		think_enabled BOOLEAN NOT NULL DEFAULT 0,
 		think_level TEXT NOT NULL DEFAULT '',
 		cloud_setting_migrated BOOLEAN NOT NULL DEFAULT 0,
@@ -264,6 +265,12 @@ func (db *database) migrate() error {
 				return fmt.Errorf("migrate v14 to v15: %w", err)
 			}
 			version = 15
+		case 15:
+			// add last_home_view column to settings table
+			if err := db.migrateV15ToV16(); err != nil {
+				return fmt.Errorf("migrate v15 to v16: %w", err)
+			}
+			version = 16
 		default:
 			// If we have a version we don't recognize, just set it to current
 			// This might happen during development
@@ -518,6 +525,21 @@ func (db *database) migrateV14ToV15() error {
 	return nil
 }

+// migrateV15ToV16 adds the last_home_view column to the settings table
+func (db *database) migrateV15ToV16() error {
+	_, err := db.conn.Exec(`ALTER TABLE settings ADD COLUMN last_home_view TEXT NOT NULL DEFAULT 'launch'`)
+	if err != nil && !duplicateColumnError(err) {
+		return fmt.Errorf("add last_home_view column: %w", err)
+	}
+
+	_, err = db.conn.Exec(`UPDATE settings SET schema_version = 16`)
+	if err != nil {
+		return fmt.Errorf("update schema version: %w", err)
+	}
+
+	return nil
+}
+
 // cleanupOrphanedData removes orphaned records that may exist due to the foreign key bug
 func (db *database) cleanupOrphanedData() error {
 	_, err := db.conn.Exec(`
@@ -1166,9 +1188,9 @@ func (db *database) getSettings() (Settings, error) {
 	var s Settings

 	err := db.conn.QueryRow(`
-		SELECT expose, survey, browser, models, agent, tools, working_dir, context_length, turbo_enabled, websearch_enabled, selected_model, sidebar_open, think_enabled, think_level, auto_update_enabled
+		SELECT expose, survey, browser, models, agent, tools, working_dir, context_length, turbo_enabled, websearch_enabled, selected_model, sidebar_open, last_home_view, think_enabled, think_level, auto_update_enabled
 		FROM settings
-	`).Scan(&s.Expose, &s.Survey, &s.Browser, &s.Models, &s.Agent, &s.Tools, &s.WorkingDir, &s.ContextLength, &s.TurboEnabled, &s.WebSearchEnabled, &s.SelectedModel, &s.SidebarOpen, &s.ThinkEnabled, &s.ThinkLevel, &s.AutoUpdateEnabled)
+	`).Scan(&s.Expose, &s.Survey, &s.Browser, &s.Models, &s.Agent, &s.Tools, &s.WorkingDir, &s.ContextLength, &s.TurboEnabled, &s.WebSearchEnabled, &s.SelectedModel, &s.SidebarOpen, &s.LastHomeView, &s.ThinkEnabled, &s.ThinkLevel, &s.AutoUpdateEnabled)
 	if err != nil {
 		return Settings{}, fmt.Errorf("get settings: %w", err)
 	}
@@ -1177,10 +1199,29 @@ func (db *database) getSettings() (Settings, error) {
 }

 func (db *database) setSettings(s Settings) error {
+	lastHomeView := strings.ToLower(strings.TrimSpace(s.LastHomeView))
+	validLaunchView := map[string]struct{}{
+		"launch":    {},
+		"openclaw":  {},
+		"claude":    {},
+		"hermes":    {},
+		"codex":     {},
+		"codex-app": {},
+		"copilot":   {},
+		"opencode":  {},
+		"droid":     {},
+		"pi":        {},
+	}
+	if lastHomeView != "chat" {
+		if _, ok := validLaunchView[lastHomeView]; !ok {
+			lastHomeView = "launch"
+		}
+	}
+
 	_, err := db.conn.Exec(`
 		UPDATE settings
-		SET expose = ?, survey = ?, browser = ?, models = ?, agent = ?, tools = ?, working_dir = ?, context_length = ?, turbo_enabled = ?, websearch_enabled = ?, selected_model = ?, sidebar_open = ?, think_enabled = ?, think_level = ?, auto_update_enabled = ?
-	`, s.Expose, s.Survey, s.Browser, s.Models, s.Agent, s.Tools, s.WorkingDir, s.ContextLength, s.TurboEnabled, s.WebSearchEnabled, s.SelectedModel, s.SidebarOpen, s.ThinkEnabled, s.ThinkLevel, s.AutoUpdateEnabled)
+		SET expose = ?, survey = ?, browser = ?, models = ?, agent = ?, tools = ?, working_dir = ?, context_length = ?, turbo_enabled = ?, websearch_enabled = ?, selected_model = ?, sidebar_open = ?, last_home_view = ?, think_enabled = ?, think_level = ?, auto_update_enabled = ?
+	`, s.Expose, s.Survey, s.Browser, s.Models, s.Agent, s.Tools, s.WorkingDir, s.ContextLength, s.TurboEnabled, s.WebSearchEnabled, s.SelectedModel, s.SidebarOpen, lastHomeView, s.ThinkEnabled, s.ThinkLevel, s.AutoUpdateEnabled)
 	if err != nil {
 		return fmt.Errorf("set settings: %w", err)
 	}
@@ -135,6 +135,45 @@ func TestMigrationV13ToV14ContextLength(t *testing.T) {
 	}
 }

+func TestMigrationV15ToV16LastHomeViewDefaultsToLaunch(t *testing.T) {
+	tmpDir := t.TempDir()
+	dbPath := filepath.Join(tmpDir, "test.db")
+
+	db, err := newDatabase(dbPath)
+	if err != nil {
+		t.Fatalf("failed to create database: %v", err)
+	}
+	defer db.Close()
+
+	if _, err := db.conn.Exec(`
+		ALTER TABLE settings DROP COLUMN last_home_view;
+		UPDATE settings SET schema_version = 15;
+	`); err != nil {
+		t.Fatalf("failed to seed v15 settings row: %v", err)
+	}
+
+	if err := db.migrate(); err != nil {
+		t.Fatalf("migration from v15 to v16 failed: %v", err)
+	}
+
+	var lastHomeView string
+	if err := db.conn.QueryRow("SELECT last_home_view FROM settings").Scan(&lastHomeView); err != nil {
+		t.Fatalf("failed to read last_home_view: %v", err)
+	}
+
+	if lastHomeView != "launch" {
+		t.Fatalf("expected last_home_view to default to launch after migration, got %q", lastHomeView)
+	}
+
+	version, err := db.getSchemaVersion()
+	if err != nil {
+		t.Fatalf("failed to get schema version: %v", err)
+	}
+	if version != currentSchemaVersion {
+		t.Fatalf("expected schema version %d, got %d", currentSchemaVersion, version)
+	}
+}
+
 func TestChatDeletionWithCascade(t *testing.T) {
 	t.Run("chat deletion cascades to related messages", func(t *testing.T) {
 		tmpDir := t.TempDir()
@@ -167,6 +167,9 @@ type Settings struct {
 	// SidebarOpen indicates if the chat sidebar is open
 	SidebarOpen bool

+	// LastHomeView stores the preferred home route target ("chat" or integration name)
+	LastHomeView string
+
 	// AutoUpdateEnabled indicates if automatic updates should be downloaded
 	AutoUpdateEnabled bool
 }
@@ -389,6 +392,10 @@ func (s *Store) Settings() (Settings, error) {
 		}
 	}

+	if settings.LastHomeView == "" {
+		settings.LastHomeView = "launch"
+	}
+
 	return settings, nil
 }

@@ -81,6 +81,62 @@ func TestStore(t *testing.T) {
 		}
 	})

+	t.Run("settings default home view is launch", func(t *testing.T) {
+		loaded, err := s.Settings()
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if loaded.LastHomeView != "launch" {
+			t.Fatalf("expected default LastHomeView to be launch, got %q", loaded.LastHomeView)
+		}
+	})
+
+	t.Run("settings empty home view falls back to launch", func(t *testing.T) {
+		if err := s.SetSettings(Settings{LastHomeView: ""}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, err := s.Settings()
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if loaded.LastHomeView != "launch" {
+			t.Fatalf("expected empty LastHomeView to fall back to launch, got %q", loaded.LastHomeView)
+		}
+	})
+
+	t.Run("settings disabled home view falls back to launch", func(t *testing.T) {
+		if err := s.SetSettings(Settings{LastHomeView: "claude-desktop"}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, err := s.Settings()
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if loaded.LastHomeView != "launch" {
+			t.Fatalf("expected disabled LastHomeView to fall back to launch, got %q", loaded.LastHomeView)
+		}
+	})
+
+	t.Run("settings codex app home view is accepted", func(t *testing.T) {
+		if err := s.SetSettings(Settings{LastHomeView: "codex-app"}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, err := s.Settings()
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if loaded.LastHomeView != "codex-app" {
+			t.Fatalf("expected codex-app LastHomeView to be preserved, got %q", loaded.LastHomeView)
+		}
+	})
+
 	t.Run("window size", func(t *testing.T) {
 		if err := s.SetWindowSize(1024, 768); err != nil {
 			t.Fatal(err)
@@ -563,6 +563,10 @@ func (b *BrowserOpen) Execute(ctx context.Context, args map[string]any) (any, st
 			return b.state.Data, pageText, nil
 		}

+		if !allowedDirectURL(ctx, url) {
+			return nil, "", fmt.Errorf("direct URL open is only allowed for URLs provided by the user")
+		}
+
 		// Page not in cache, need to crawl it
 		if b.crawlPage == nil {
 			b.crawlPage = &BrowserCrawler{}
@@ -65,6 +65,27 @@ func TestBrowserOpen_UseCacheByURL(t *testing.T) {
 	}
 }

+func TestBrowserOpen_RejectsUncachedDirectURL(t *testing.T) {
+	b := NewBrowser(&responses.BrowserStateData{PageStack: []string{}, ViewTokens: 1024, URLToPage: map[string]*responses.Page{}})
+	bo := NewBrowserOpen(b)
+
+	_, _, err := bo.Execute(t.Context(), map[string]any{"id": "https://attacker.example/?data=secret"})
+	if err == nil || !strings.Contains(err.Error(), "only allowed for URLs provided by the user") {
+		t.Fatalf("expected direct URL rejection, got %v", err)
+	}
+}
+
+func TestDirectURLsFromText_AllowsExactUserURLsOnly(t *testing.T) {
+	ctx := WithAllowedDirectURLs(t.Context(), "summarize https://example.com/article?q=1 please")
+
+	if !allowedDirectURL(ctx, "https://example.com/article?q=1") {
+		t.Fatal("expected exact user-provided URL to be allowed")
+	}
+	if allowedDirectURL(ctx, "https://example.com/article?q=secret") {
+		t.Fatal("did not expect modified URL to be allowed")
+	}
+}
+
 func TestDisplayPage_InvalidLoc(t *testing.T) {
 	b := NewBrowser(&responses.BrowserStateData{PageStack: []string{}, ViewTokens: 1024, URLToPage: map[string]*responses.Page{}})
 	p := makeTestPage("https://example.com/x")
@@ -0,0 +1,61 @@
+//go:build windows || darwin
+
+package tools
+
+import (
+	"context"
+	"regexp"
+	"strings"
+)
+
+type directURLContextKey struct{}
+
+var directURLPattern = regexp.MustCompile("https?://[^\\s<>\"'`]+")
+
+func WithAllowedDirectURLs(ctx context.Context, text string) context.Context {
+	allowed := make(map[string]struct{})
+	for _, match := range directURLPattern.FindAllString(text, -1) {
+		addAllowedDirectURLToMap(allowed, match)
+	}
+	return context.WithValue(ctx, directURLContextKey{}, allowed)
+}
+
+func addAllowedDirectURL(ctx context.Context, raw string) {
+	allowed, _ := ctx.Value(directURLContextKey{}).(map[string]struct{})
+	addAllowedDirectURLToMap(allowed, raw)
+}
+
+func addAllowedDirectURLToMap(allowed map[string]struct{}, raw string) {
+	if allowed == nil {
+		return
+	}
+
+	raw = cleanDirectURL(raw)
+	if raw == "" {
+		return
+	}
+
+	allowed[raw] = struct{}{}
+}
+
+func allowedDirectURL(ctx context.Context, raw string) bool {
+	allowed, _ := ctx.Value(directURLContextKey{}).(map[string]struct{})
+	cleaned := cleanDirectURL(raw)
+	if cleaned == "" || cleaned != raw {
+		return false
+	}
+
+	_, ok := allowed[cleaned]
+	return ok
+}
+
+func cleanDirectURL(raw string) string {
+	raw = strings.TrimSpace(raw)
+	raw = strings.TrimRight(raw, ".,;:!?)]}")
+
+	if !strings.HasPrefix(raw, "http://") && !strings.HasPrefix(raw, "https://") {
+		return ""
+	}
+
+	return raw
+}
@@ -0,0 +1,21 @@
+//go:build windows || darwin
+
+package tools
+
+import "testing"
+
+func TestDirectURLsFromText_RejectsChangedToolArgument(t *testing.T) {
+	ctx := WithAllowedDirectURLs(t.Context(), "summarize https://attacker.example/x")
+
+	if allowedDirectURL(ctx, "https://attacker.example/x!!!!") {
+		t.Fatal("expected changed tool argument to be rejected")
+	}
+}
+
+func TestDirectURLsFromText_ExtractsMarkdownCodeSpanURL(t *testing.T) {
+	ctx := WithAllowedDirectURLs(t.Context(), "summarize `https://example.com/privacy`")
+
+	if !allowedDirectURL(ctx, "https://example.com/privacy") {
+		t.Fatal("expected URL wrapped in backticks to be allowed")
+	}
+}
@@ -67,11 +67,17 @@ func (w *WebFetch) Execute(ctx context.Context, args map[string]any) (any, strin
 	if !ok || strings.TrimSpace(urlStr) == "" {
 		return nil, "", fmt.Errorf("url must be a non-empty string")
 	}
+	if !allowedDirectURL(ctx, urlStr) {
+		return nil, "", fmt.Errorf("web fetch is only allowed for URLs provided by the user")
+	}

 	result, err := performWebFetch(ctx, urlStr)
 	if err != nil {
 		return nil, "", err
 	}
+	for _, link := range result.Links {
+		addAllowedDirectURL(ctx, link)
+	}

 	return result, "", nil
 }
@@ -88,6 +88,9 @@ func (w *WebSearch) Execute(ctx context.Context, args map[string]any) (any, stri
 	if err != nil {
 		return nil, "", err
 	}
+	for _, result := range result.Results {
+		addAllowedDirectURL(ctx, result.URL)
+	}

 	return result, "", nil
 }
@@ -414,6 +414,7 @@ export class Settings {
    ThinkLevel: string;
    SelectedModel: string;
    SidebarOpen: boolean;
+    LastHomeView: string;
    AutoUpdateEnabled: boolean;

    constructor(source: any = {}) {
@@ -432,6 +433,7 @@ export class Settings {
        this.ThinkLevel = source["ThinkLevel"];
        this.SelectedModel = source["SelectedModel"];
        this.SidebarOpen = source["SidebarOpen"];
+        this.LastHomeView = source["LastHomeView"];
        this.AutoUpdateEnabled = source["AutoUpdateEnabled"];
    }
 }
@@ -0,0 +1 @@
+<svg height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Claude Code</title><path clip-rule="evenodd" d="M20.998 10.949H24v3.102h-3v3.028h-1.487V20H18v-2.921h-1.487V20H15v-2.921H9V20H7.488v-2.921H6V20H4.487v-2.921H3V14.05H0V10.95h3V5h17.998v5.949zM6 10.949h1.488V8.102H6v2.847zm10.51 0H18V8.102h-1.49v2.847z" fill="#D97757" fill-rule="evenodd"></path></svg>
@@ -0,0 +1,7 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!-- Generated by Pixelmator Pro 3.6.17 -->
+<svg width="1200" height="1200" viewBox="0 0 1200 1200" xmlns="http://www.w3.org/2000/svg">
+    <g id="g314">
+        <path id="path147" fill="#d97757" stroke="none" d="M 233.959793 800.214905 L 468.644287 668.536987 L 472.590637 657.100647 L 468.644287 650.738403 L 457.208069 650.738403 L 417.986633 648.322144 L 283.892639 644.69812 L 167.597321 639.865845 L 54.926208 633.825623 L 26.577238 627.785339 L 3.3e-05 592.751709 L 2.73832 575.27533 L 26.577238 559.248352 L 60.724873 562.228149 L 136.187973 567.382629 L 249.422867 575.194763 L 331.570496 580.026978 L 453.261841 592.671082 L 472.590637 592.671082 L 475.328857 584.859009 L 468.724915 580.026978 L 463.570557 575.194763 L 346.389313 495.785217 L 219.543671 411.865906 L 153.100723 363.543762 L 117.181267 339.060425 L 99.060455 316.107361 L 91.248367 266.01355 L 123.865784 230.093994 L 167.677887 233.073853 L 178.872513 236.053772 L 223.248367 270.201477 L 318.040283 343.570496 L 441.825592 434.738342 L 459.946411 449.798706 L 467.194672 444.64447 L 468.080597 441.020203 L 459.946411 427.409485 L 392.617493 305.718323 L 320.778564 181.932983 L 288.80542 130.630859 L 280.348999 99.865845 C 277.369171 87.221436 275.194641 76.590698 275.194641 63.624268 L 312.322174 13.20813 L 332.8591 6.604126 L 382.389313 13.20813 L 403.248352 31.328979 L 434.013519 101.71814 L 483.865753 212.537048 L 561.181274 363.221497 L 583.812134 407.919434 L 595.892639 449.315491 L 600.40271 461.959839 L 608.214783 461.959839 L 608.214783 454.711609 L 614.577271 369.825623 L 626.335632 265.61084 L 637.771851 131.516846 L 641.718201 93.745117 L 660.402832 48.483276 L 697.530334 24.000122 L 726.52356 37.852417 L 750.362549 72 L 747.060486 94.067139 L 732.886047 186.201416 L 705.100708 330.52356 L 686.979919 427.167847 L 697.530334 427.167847 L 709.61084 415.087341 L 758.496704 350.174561 L 840.644348 247.490051 L 876.885925 206.738342 L 919.167847 161.71814 L 946.308838 140.29541 L 997.61084 140.29541 L 1035.38269 196.429626 L 1018.469849 254.416199 L 965.637634 321.422852 L 921.825562 378.201538 L 859.006714 462.765259 L 819.785278 530.41626 L 823.409424 535.812073 L 832.75177 534.92627 L 974.657776 504.724915 L 1051.328979 490.872559 L 1142.818848 475.167786 L 1184.214844 494.496582 L 1188.724854 514.147644 L 1172.456421 554.335693 L 1074.604126 578.496765 L 959.838989 601.449829 L 788.939636 641.879272 L 786.845764 643.409485 L 789.261841 646.389343 L 866.255127 653.637634 L 899.194702 655.409424 L 979.812134 655.409424 L 1129.932861 666.604187 L 1169.154419 692.537109 L 1192.671265 724.268677 L 1188.724854 748.429688 L 1128.322144 779.194641 L 1046.818848 759.865845 L 856.590759 714.604126 L 791.355774 698.335754 L 782.335693 698.335754 L 782.335693 703.731567 L 836.69812 756.885986 L 936.322205 846.845581 L 1061.073975 962.81897 L 1067.436279 991.490112 L 1051.409424 1014.120911 L 1034.496704 1011.704712 L 924.885986 929.234924 L 882.604126 892.107544 L 786.845764 811.48999 L 780.483276 811.48999 L 780.483276 819.946289 L 802.550415 852.241699 L 919.087341 1027.409424 L 925.127625 1081.127686 L 916.671204 1098.604126 L 886.469849 1109.154419 L 853.288696 1103.114136 L 785.073914 1007.355835 L 714.684631 899.516785 L 657.906067 802.872498 L 650.979858 806.81897 L 617.476624 1167.704834 L 601.771851 1186.147705 L 565.530212 1200 L 535.328857 1177.046997 L 519.302124 1139.919556 L 535.328857 1066.550537 L 554.657776 970.792053 L 570.362488 894.68457 L 584.536926 800.134277 L 592.993347 768.724976 L 592.429626 766.630859 L 585.503479 767.516968 L 514.22821 865.369263 L 405.825531 1011.865906 L 320.053711 1103.677979 L 299.516815 1111.812256 L 263.919525 1093.369263 L 267.221497 1060.429688 L 287.114136 1031.114136 L 405.825531 880.107361 L 477.422913 786.52356 L 523.651062 732.483276 L 523.328918 724.671265 L 520.590698 724.671265 L 205.288605 929.395935 L 149.154434 936.644409 L 124.993355 914.01355 L 127.973183 876.885986 L 139.409409 864.80542 L 234.201385 799.570435 L 233.879227 799.8927 Z"/>
+    </g>
+</svg>
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 320 320"><path fill="#fff" d="m297.06 130.97c7.26-21.79 4.76-45.66-6.85-65.48-17.46-30.4-52.56-46.04-86.84-38.68-15.25-17.18-37.16-26.95-60.13-26.81-35.04-.08-66.13 22.48-76.91 55.82-22.51 4.61-41.94 18.7-53.31 38.67-17.59 30.32-13.58 68.54 9.92 94.54-7.26 21.79-4.76 45.66 6.85 65.48 17.46 30.4 52.56 46.04 86.84 38.68 15.24 17.18 37.16 26.95 60.13 26.8 35.06.09 66.16-22.49 76.94-55.86 22.51-4.61 41.94-18.7 53.31-38.67 17.57-30.32 13.55-68.51-9.94-94.51zm-120.28 168.11c-14.03.02-27.62-4.89-38.39-13.88.49-.26 1.34-.73 1.89-1.07l63.72-36.8c3.26-1.85 5.26-5.32 5.24-9.07v-89.83l26.93 15.55c.29.14.48.42.52.74v74.39c-.04 33.08-26.83 59.9-59.91 59.97zm-128.84-55.03c-7.03-12.14-9.56-26.37-7.15-40.18.47.28 1.3.79 1.89 1.13l63.72 36.8c3.23 1.89 7.23 1.89 10.47 0l77.79-44.92v31.1c.02.32-.13.63-.38.83l-64.41 37.19c-28.69 16.52-65.33 6.7-81.92-21.95zm-16.77-139.09c7-12.16 18.05-21.46 31.21-26.29 0 .55-.03 1.52-.03 2.2v73.61c-.02 3.74 1.98 7.21 5.23 9.06l77.79 44.91-26.93 15.55c-.27.18-.61.21-.91.08l-64.42-37.22c-28.63-16.58-38.45-53.21-21.95-81.89zm221.26 51.49-77.79-44.92 26.93-15.54c.27-.18.61-.21.91-.08l64.42 37.19c28.68 16.57 38.51 53.26 21.94 81.94-7.01 12.14-18.05 21.44-31.2 26.28v-75.81c.03-3.74-1.96-7.2-5.2-9.06zm26.8-40.34c-.47-.29-1.3-.79-1.89-1.13l-63.72-36.8c-3.23-1.89-7.23-1.89-10.47 0l-77.79 44.92v-31.1c-.02-.32.13-.63.38-.83l64.41-37.16c28.69-16.55 65.37-6.7 81.91 22 6.99 12.12 9.52 26.31 7.15 40.1zm-168.51 55.43-26.94-15.55c-.29-.14-.48-.42-.52-.74v-74.39c.02-33.12 26.89-59.96 60.01-59.94 14.01 0 27.57 4.92 38.34 13.88-.49.26-1.33.73-1.89 1.07l-63.72 36.8c-3.26 1.85-5.26 5.31-5.24 9.06l-.04 89.79zm14.63-31.54 34.65-20.01 34.65 20v40.01l-34.65 20-34.65-20z"/></svg>
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 320 320"><path d="m297.06 130.97c7.26-21.79 4.76-45.66-6.85-65.48-17.46-30.4-52.56-46.04-86.84-38.68-15.25-17.18-37.16-26.95-60.13-26.81-35.04-.08-66.13 22.48-76.91 55.82-22.51 4.61-41.94 18.7-53.31 38.67-17.59 30.32-13.58 68.54 9.92 94.54-7.26 21.79-4.76 45.66 6.85 65.48 17.46 30.4 52.56 46.04 86.84 38.68 15.24 17.18 37.16 26.95 60.13 26.8 35.06.09 66.16-22.49 76.94-55.86 22.51-4.61 41.94-18.7 53.31-38.67 17.57-30.32 13.55-68.51-9.94-94.51zm-120.28 168.11c-14.03.02-27.62-4.89-38.39-13.88.49-.26 1.34-.73 1.89-1.07l63.72-36.8c3.26-1.85 5.26-5.32 5.24-9.07v-89.83l26.93 15.55c.29.14.48.42.52.74v74.39c-.04 33.08-26.83 59.9-59.91 59.97zm-128.84-55.03c-7.03-12.14-9.56-26.37-7.15-40.18.47.28 1.3.79 1.89 1.13l63.72 36.8c3.23 1.89 7.23 1.89 10.47 0l77.79-44.92v31.1c.02.32-.13.63-.38.83l-64.41 37.19c-28.69 16.52-65.33 6.7-81.92-21.95zm-16.77-139.09c7-12.16 18.05-21.46 31.21-26.29 0 .55-.03 1.52-.03 2.2v73.61c-.02 3.74 1.98 7.21 5.23 9.06l77.79 44.91-26.93 15.55c-.27.18-.61.21-.91.08l-64.42-37.22c-28.63-16.58-38.45-53.21-21.95-81.89zm221.26 51.49-77.79-44.92 26.93-15.54c.27-.18.61-.21.91-.08l64.42 37.19c28.68 16.57 38.51 53.26 21.94 81.94-7.01 12.14-18.05 21.44-31.2 26.28v-75.81c.03-3.74-1.96-7.2-5.2-9.06zm26.8-40.34c-.47-.29-1.3-.79-1.89-1.13l-63.72-36.8c-3.23-1.89-7.23-1.89-10.47 0l-77.79 44.92v-31.1c-.02-.32.13-.63.38-.83l64.41-37.16c28.69-16.55 65.37-6.7 81.91 22 6.99 12.12 9.52 26.31 7.15 40.1zm-168.51 55.43-26.94-15.55c-.29-.14-.48-.42-.52-.74v-74.39c.02-33.12 26.89-59.96 60.01-59.94 14.01 0 27.57 4.92 38.34 13.88-.49.26-1.33.73-1.89 1.07l-63.72 36.8c-3.26 1.85-5.26 5.31-5.24 9.06l-.04 89.79zm14.63-31.54 34.65-20.01 34.65 20v40.01l-34.65 20-34.65-20z"/></svg>
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" fill="currentColor" fill-rule="evenodd" style="flex:none;line-height:1" viewBox="0 2.5 24 19"><path d="M19.245 5.364c1.322 1.36 1.877 3.216 2.11 5.817.622 0 1.2.135 1.592.654l.73.964c.21.278.323.61.323.955v2.62c0 .339-.173.669-.453.868C20.239 19.602 16.157 21.5 12 21.5c-4.6 0-9.205-2.583-11.547-4.258-.28-.2-.452-.53-.453-.868v-2.62c0-.345.113-.679.321-.956l.73-.963c.392-.517.974-.654 1.593-.654l.029-.297c.25-2.446.81-4.213 2.082-5.52 2.461-2.54 5.71-2.851 7.146-2.864h.198c1.436.013 4.685.323 7.146 2.864zm-7.244 4.328c-.284 0-.613.016-.962.05-.123.447-.305.85-.57 1.108-1.05 1.023-2.316 1.18-2.994 1.18-.638 0-1.306-.13-1.851-.464-.516.165-1.012.403-1.044.996a65.882 65.882 0 00-.063 2.884l-.002.48c-.002.563-.005 1.126-.013 1.69.002.326.204.63.51.765 2.482 1.102 4.83 1.657 6.99 1.657 2.156 0 4.504-.555 6.985-1.657a.854.854 0 00.51-.766c.03-1.682.006-3.372-.076-5.053-.031-.596-.528-.83-1.046-.996-.546.333-1.212.464-1.85.464-.677 0-1.942-.157-2.993-1.18-.266-.258-.447-.661-.57-1.108-.32-.032-.64-.049-.96-.05zm-2.525 4.013c.539 0 .976.426.976.95v1.753c0 .525-.437.95-.976.95a.964.964 0 01-.976-.95v-1.752c0-.525.437-.951.976-.951zm5 0c.539 0 .976.426.976.95v1.753c0 .525-.437.95-.976.95a.964.964 0 01-.976-.95v-1.752c0-.525.437-.951.976-.951zM7.635 5.087c-1.05.102-1.935.438-2.385.906-.975 1.037-.765 3.668-.21 4.224.405.394 1.17.657 1.995.657h.09c.649-.013 1.785-.176 2.73-1.11.435-.41.705-1.433.675-2.47-.03-.834-.27-1.52-.63-1.813-.39-.336-1.275-.482-2.265-.394zm6.465.394c-.36.292-.6.98-.63 1.813-.03 1.037.24 2.06.675 2.47.968.957 2.136 1.104 2.776 1.11h.044c.825 0 1.59-.263 1.995-.657.555-.556.765-3.187-.21-4.224-.45-.468-1.335-.804-2.385-.906-.99-.088-1.875.058-2.265.394zM12 7.615c-.24 0-.525.015-.84.044.03.16.045.336.06.526l-.001.159a2.94 2.94 0 01-.014.25c.225-.022.425-.027.612-.028h.366c.187 0 .387.006.612.028-.015-.146-.015-.277-.015-.409.015-.19.03-.365.06-.526a9.29 9.29 0 00-.84-.044z" fill="white"/></svg>
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" fill="currentColor" fill-rule="evenodd" style="flex:none;line-height:1" viewBox="0 2.5 24 19"><path d="M19.245 5.364c1.322 1.36 1.877 3.216 2.11 5.817.622 0 1.2.135 1.592.654l.73.964c.21.278.323.61.323.955v2.62c0 .339-.173.669-.453.868C20.239 19.602 16.157 21.5 12 21.5c-4.6 0-9.205-2.583-11.547-4.258-.28-.2-.452-.53-.453-.868v-2.62c0-.345.113-.679.321-.956l.73-.963c.392-.517.974-.654 1.593-.654l.029-.297c.25-2.446.81-4.213 2.082-5.52 2.461-2.54 5.71-2.851 7.146-2.864h.198c1.436.013 4.685.323 7.146 2.864zm-7.244 4.328c-.284 0-.613.016-.962.05-.123.447-.305.85-.57 1.108-1.05 1.023-2.316 1.18-2.994 1.18-.638 0-1.306-.13-1.851-.464-.516.165-1.012.403-1.044.996a65.882 65.882 0 00-.063 2.884l-.002.48c-.002.563-.005 1.126-.013 1.69.002.326.204.63.51.765 2.482 1.102 4.83 1.657 6.99 1.657 2.156 0 4.504-.555 6.985-1.657a.854.854 0 00.51-.766c.03-1.682.006-3.372-.076-5.053-.031-.596-.528-.83-1.046-.996-.546.333-1.212.464-1.85.464-.677 0-1.942-.157-2.993-1.18-.266-.258-.447-.661-.57-1.108-.32-.032-.64-.049-.96-.05zm-2.525 4.013c.539 0 .976.426.976.95v1.753c0 .525-.437.95-.976.95a.964.964 0 01-.976-.95v-1.752c0-.525.437-.951.976-.951zm5 0c.539 0 .976.426.976.95v1.753c0 .525-.437.95-.976.95a.964.964 0 01-.976-.95v-1.752c0-.525.437-.951.976-.951zM7.635 5.087c-1.05.102-1.935.438-2.385.906-.975 1.037-.765 3.668-.21 4.224.405.394 1.17.657 1.995.657h.09c.649-.013 1.785-.176 2.73-1.11.435-.41.705-1.433.675-2.47-.03-.834-.27-1.52-.63-1.813-.39-.336-1.275-.482-2.265-.394zm6.465.394c-.36.292-.6.98-.63 1.813-.03 1.037.24 2.06.675 2.47.968.957 2.136 1.104 2.776 1.11h.044c.825 0 1.59-.263 1.995-.657.555-.556.765-3.187-.21-4.224-.45-.468-1.335-.804-2.385-.906-.99-.088-1.875.058-2.265.394zM12 7.615c-.24 0-.525.015-.84.044.03.16.045.336.06.526l-.001.159a2.94 2.94 0 01-.014.25c.225-.022.425-.027.612-.028h.366c.187 0 .387.006.612.028-.015-.146-.015-.277-.015-.409.015-.19.03-.365.06-.526a9.29 9.29 0 00-.84-.044z"/></svg>
@@ -0,0 +1,181 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<svg xmlns="http://www.w3.org/2000/svg" width="1000" height="1000" viewBox="0 0 1000 1000"><circle cx="500.0" cy="500.0" r="500.0" fill="white"/><g transform="translate(100.0 100.0) scale(0.8333333333333334)"><g transform="translate(0.000000,960.000000) scale(0.100000,-0.100000)"
+fill="black" stroke="none">
+<path d="M4485 9589 c-248 -27 -432 -60 -730 -130 -458 -108 -798 -230 -1207
+-435 -533 -267 -1072 -675 -1358 -1030 -205 -255 -442 -748 -535 -1114 -108
+-426 -97 -870 29 -1160 73 -169 236 -369 381 -467 139 -94 425 -206 425 -167
+0 3 -26 32 -58 63 -74 71 -147 182 -184 278 -16 41 -27 77 -25 79 2 3 31 -28
+63 -68 91 -114 153 -177 231 -235 40 -29 77 -62 83 -73 7 -13 4 -85 -10 -242
+-10 -123 -24 -281 -30 -353 -6 -71 -29 -296 -51 -500 -135 -1262 -202 -1568
+-378 -1733 -69 -64 -105 -77 -216 -77 -132 0 -188 20 -271 97 -110 103 -165
+248 -167 438 -1 123 12 191 60 318 20 51 33 95 29 99 -10 10 -92 -74 -136
+-137 -153 -223 -204 -504 -146 -790 19 -91 97 -252 161 -333 112 -141 316
+-237 504 -237 180 0 419 118 591 290 81 81 107 115 196 255 50 78 54 56 13
+-71 -107 -337 -343 -577 -613 -625 -213 -39 -544 99 -707 295 -93 111 -133
+199 -173 381 -34 153 -34 154 -46 135 -15 -23 -12 -194 5 -305 42 -280 149
+-488 327 -636 133 -111 287 -195 465 -254 62 -20 115 -40 117 -44 3 -4 12 -43
+21 -87 44 -222 199 -385 416 -438 96 -24 265 -21 359 5 138 38 281 148 388
+297 39 55 48 63 50 45 4 -27 -38 -185 -78 -294 -80 -217 -176 -374 -312 -515
+-54 -55 -98 -104 -98 -107 0 -4 245 -7 545 -7 l544 0 126 66 c69 36 130 64
+135 62 6 -2 -10 -28 -34 -59 -46 -59 -48 -69 -10 -69 17 0 32 15 58 59 36 59
+51 71 87 71 23 0 25 -27 4 -77 -8 -19 -15 -39 -15 -44 0 -12 447 -11 470 1 10
+5 85 87 168 182 257 297 400 415 317 263 -16 -30 -26 -57 -23 -60 11 -12 210
+100 383 215 115 76 226 143 375 225 58 32 178 100 266 151 150 87 244 132 244
+117 0 -4 -78 -86 -174 -182 -207 -208 -246 -254 -334 -386 -48 -71 -122 -157
+-259 -298 -117 -120 -193 -206 -193 -217 0 -19 8 -20 115 -20 101 0 116 2 122
+18 19 54 112 249 156 327 124 221 283 436 369 502 87 67 225 152 337 209 103
+51 297 117 384 130 38 6 29 -2 -92 -75 -74 -45 -170 -107 -214 -139 -106 -76
+-247 -225 -332 -351 -67 -99 -70 -102 -79 -77 -6 14 -13 26 -18 26 -11 0 -57
+-75 -94 -155 -63 -136 -124 -376 -103 -401 15 -18 152 -19 166 -2 6 7 18 36
+27 63 23 72 66 131 217 301 302 339 430 464 590 577 148 104 233 128 520 148
+205 14 255 9 447 -38 118 -29 183 -65 298 -163 229 -195 538 -577 579 -715 24
+-83 71 -117 109 -79 9 8 16 27 16 42 0 62 -40 212 -73 277 -158 313 -551 635
+-922 755 -127 41 -142 55 -52 46 117 -11 231 -35 321 -65 158 -54 259 -123
+405 -277 164 -174 283 -379 322 -556 11 -51 25 -113 32 -137 36 -128 148 -81
+117 49 -14 60 -7 106 27 177 36 74 62 94 224 179 254 133 425 260 561 417 272
+315 403 732 358 1138 -24 218 -84 401 -179 545 -65 98 -155 203 -166 192 -3
+-3 7 -37 23 -75 136 -309 136 -725 2 -1063 -130 -330 -441 -760 -633 -879 -60
+-37 -60 -20 2 54 478 577 578 1337 315 2390 -25 99 -86 326 -136 505 -169 611
+-386 1539 -494 2109 -161 857 -200 998 -442 1606 -161 405 -321 692 -529 950
+-93 114 -322 343 -433 431 -203 160 -497 332 -737 428 -293 118 -702 220 -978
+246 -129 12 -406 11 -520 -1z m-267 -214 c301 -47 596 -191 852 -419 85 -76
+266 -270 345 -371 294 -376 520 -873 640 -1409 45 -199 45 -222 3 -244 -18 -9
+-45 -30 -61 -45 -25 -23 -28 -33 -23 -60 9 -44 21 -54 121 -98 91 -40 225
+-124 225 -140 0 -9 -33 5 -211 89 -126 60 -159 94 -167 173 -5 50 14 89 43 89
+26 0 65 49 65 81 0 52 -22 110 -48 127 -23 15 -37 14 -186 -7 -172 -25 -210
+-36 -240 -69 -23 -27 -31 -99 -14 -131 16 -29 30 -37 85 -51 26 -7 44 -17 48
+-29 10 -32 -15 -153 -41 -197 -27 -49 -131 -161 -139 -152 -9 9 36 77 85 126
+23 24 48 60 56 79 13 31 13 38 -3 71 -10 20 -29 42 -43 48 -14 7 -37 18 -52
+24 -33 15 -42 39 -63 155 -24 133 -90 393 -135 532 -119 367 -287 686 -483
+919 -175 208 -434 427 -659 557 -229 131 -498 197 -804 197 -136 0 -185 -4
+-313 -26 -87 -14 -31 14 137 70 186 61 289 89 437 117 120 23 376 20 543 -6z
+m2913 -962 c50 -53 113 -138 198 -268 76 -116 60 -104 -41 32 -34 46 -63 81
+-66 79 -2 -2 2 -30 9 -63 9 -44 9 -86 1 -171 -6 -63 -13 -116 -16 -119 -3 -4
+-15 2 -26 12 -32 29 -34 12 -5 -57 50 -118 66 -160 62 -164 -2 -2 -29 34 -61
+81 -31 47 -61 83 -66 80 -6 -4 -7 -24 -4 -47 7 -39 6 -40 -14 -27 -12 7 -25
+10 -29 5 -17 -17 -4 -106 31 -210 20 -60 35 -111 33 -112 -2 -2 -30 43 -62 99
+-60 102 -173 233 -182 210 -2 -7 26 -82 62 -168 37 -85 65 -160 63 -166 -2 -6
+-43 66 -91 160 -63 121 -93 171 -106 171 -9 0 -37 -21 -61 -46 -55 -56 -56
+-56 -248 20 -78 31 -156 59 -172 63 l-30 6 24 -54 c13 -30 51 -114 85 -188 34
+-73 60 -135 58 -137 -2 -3 -21 24 -42 58 -56 91 -85 128 -102 128 -13 0 -14
+-8 -9 -42 l7 -42 -83 80 c-109 106 -163 133 -260 126 -35 -3 -38 0 -53 34 -9
+21 -13 44 -11 51 7 17 58 16 147 -2 101 -21 104 -17 92 106 -6 52 -7 98 -3
+102 4 5 25 -20 47 -55 22 -35 47 -68 54 -75 18 -14 278 -83 316 -83 23 0 42
+13 86 59 61 64 64 72 42 111 -19 33 -19 54 0 46 11 -4 23 6 37 30 l21 35 66
+-3 66 -3 -2 29 c-1 16 -25 63 -52 105 -28 41 -49 77 -47 78 2 2 47 -42 101
+-97 75 -76 105 -100 125 -100 14 0 35 -7 47 -15 20 -14 22 -14 27 2 3 10 6 72
+7 138 2 102 -1 128 -19 175 -12 30 -22 56 -22 58 0 9 40 -22 71 -55z m-2642
+-139 c29 -35 60 -64 67 -64 8 0 52 27 97 61 l82 60 47 -51 c59 -63 161 -218
+218 -327 23 -46 48 -83 55 -83 8 0 26 5 41 11 49 18 73 4 104 -64 33 -70 48
+-127 33 -127 -6 0 -32 9 -58 21 -28 12 -56 18 -71 15 -21 -6 -27 1 -55 56 -67
+132 -208 340 -244 361 -10 5 -19 -1 -29 -22 -29 -55 -35 -106 -20 -183 8 -40
+14 -74 14 -75 0 -1 -12 2 -26 8 l-27 10 7 -62 c6 -61 6 -62 -14 -44 -15 14
+-31 17 -72 13 -58 -6 -88 -32 -88 -76 l0 -26 -29 34 c-29 35 -30 35 -122 38
+-52 2 -99 8 -106 14 -15 12 -83 132 -83 146 0 6 23 39 50 73 56 70 68 99 52
+134 -12 27 -15 25 81 46 65 14 68 21 42 105 -26 85 -18 85 54 -2z m-397 -294
+c84 -126 196 -352 237 -475 33 -99 28 -115 -23 -66 -45 44 -55 29 -49 -77 5
+-95 -4 -97 -33 -7 -22 67 -52 125 -64 125 -6 0 -10 -39 -11 -92 0 -51 -4 -101
+-8 -111 -9 -23 -10 -22 -85 101 -32 50 -62 92 -68 92 -7 0 -9 -22 -4 -70 6
+-74 -3 -90 -24 -40 -21 50 -33 54 -87 30 -26 -11 -57 -30 -69 -41 -20 -19 -21
+-19 -75 10 -30 17 -91 41 -137 54 -46 13 -89 30 -97 38 -16 16 -45 142 -36
+157 3 5 58 33 121 61 l114 51 21 -26 21 -26 49 39 c51 40 69 66 80 116 5 23
+10 28 32 25 19 -2 28 -11 37 -38 37 -115 42 114 6 250 -45 171 -47 164 22 90
+33 -36 92 -112 130 -170z m-1789 73 c-3 -10 -32 -77 -63 -148 -48 -109 -137
+-340 -242 -628 -11 -32 -22 -56 -24 -54 -2 2 -9 32 -14 68 -24 141 -20 131
+-47 124 -13 -3 -50 -18 -81 -33 -43 -20 -62 -37 -79 -67 -32 -59 -40 -65 -88
+-65 -55 0 -56 6 -16 106 29 75 176 339 194 351 12 7 14 14 -47 -125 -25 -56
+-46 -113 -46 -127 0 -25 0 -25 35 -11 109 46 179 120 301 316 135 217 241 360
+217 293z m-869 -35 c-4 -7 -33 -49 -64 -93 -140 -200 -268 -431 -368 -665
+-114 -268 -153 -314 -74 -86 41 115 44 129 29 137 -28 16 -39 80 -22 128 27
+77 98 202 139 246 58 61 355 345 362 345 3 0 2 -6 -2 -12z m1415 -533 l1 -130
+-23 39 c-26 47 -41 51 -45 14 -4 -36 -18 -35 -37 1 -8 17 -19 32 -25 36 -5 3
+-60 -10 -122 -30 -76 -25 -130 -36 -165 -36 -29 1 -82 -5 -118 -14 -99 -23
+-109 -21 -149 29 -20 23 -36 50 -36 58 0 12 55 182 75 231 11 27 38 21 147
+-34 76 -38 108 -49 130 -45 21 4 28 2 28 -9 0 -10 11 -15 34 -15 45 0 153 34
+177 56 10 9 35 69 55 133 56 180 58 180 65 1 4 -85 7 -213 8 -285z m-1394 272
+c-31 -55 -32 -83 -2 -91 47 -12 65 -6 97 34 18 23 34 39 36 38 2 -2 -20 -48
+-48 -102 l-50 -99 33 7 c84 17 149 19 149 5 0 -8 -37 -104 -82 -213 -74 -177
+-83 -195 -86 -162 -4 50 -26 56 -66 17 -17 -17 -35 -31 -39 -31 -4 0 -7 18 -7
+40 0 28 -6 43 -18 51 -15 10 -18 21 -14 65 8 93 -18 69 -89 -80 -35 -74 -65
+-133 -67 -131 -3 2 5 35 17 74 11 38 21 74 21 80 0 6 -35 11 -87 13 l-88 3 3
+30 c4 42 162 364 187 381 11 8 44 14 76 14 55 0 59 2 93 42 20 23 36 45 36 50
+0 4 5 8 10 8 6 0 -1 -20 -15 -43z m1710 6 c120 -17 143 -19 164 -12 10 3 21
+-17 37 -67 24 -77 75 -306 69 -312 -2 -2 -21 14 -43 37 l-39 41 -145 0 c-128
+0 -148 2 -170 19 -16 13 -29 17 -39 10 -8 -5 -17 -9 -20 -9 -10 0 -66 178 -74
+232 -4 25 -4 56 0 68 7 21 11 22 79 16 39 -4 121 -14 181 -23z m-79 -988 c74
+-190 129 -303 247 -514 76 -136 79 -145 62 -157 -24 -17 -76 -18 -98 -1 -21
+16 -126 237 -165 347 -30 87 -143 516 -141 541 1 19 17 -19 95 -216z m3644 61
+c64 -163 185 -534 301 -926 66 -223 147 -490 179 -595 175 -566 250 -858 321
+-1240 22 -121 43 -231 46 -245 4 -18 3 -22 -6 -15 -6 6 -22 71 -36 145 -58
+313 -100 487 -200 820 -40 135 -94 317 -120 405 -161 550 -413 1387 -465 1540
+-55 165 -67 205 -56 194 2 -2 18 -39 36 -83z m-4395 -812 c61 -60 129 -118
+149 -128 46 -22 69 -19 238 25 70 18 130 30 133 27 3 -3 -19 -42 -49 -87 -53
+-78 -75 -125 -63 -137 14 -14 111 32 205 96 57 38 136 85 176 104 84 40 299
+116 327 116 12 0 37 -24 66 -65 25 -36 54 -67 62 -69 9 -2 178 -1 376 2 l360
+7 10 70 c10 68 10 69 22 40 7 -16 17 -49 23 -73 14 -53 18 -55 187 -82 178
+-28 191 -33 200 -78 12 -57 8 -612 -6 -742 -17 -170 -53 -395 -141 -880 -206
+-1142 -248 -1540 -194 -1863 8 -49 12 -97 9 -107 -8 -24 -195 -200 -213 -200
+-7 0 -21 14 -30 31 -140 256 -353 528 -467 594 -66 39 -95 39 -361 1 -137 -19
+-303 -40 -368 -47 -128 -12 -307 -7 -368 11 -54 16 -140 78 -185 132 -41 51
+-348 610 -438 801 -105 220 -178 478 -191 667 -7 97 11 328 25 343 5 5 23 -17
+41 -49 18 -32 72 -97 125 -148 54 -53 97 -104 101 -120 12 -50 -1 -119 -34
+-170 -22 -33 -32 -62 -32 -88 0 -45 31 -126 70 -182 23 -35 28 -50 23 -82 -3
+-24 4 -70 17 -119 18 -68 28 -87 68 -128 59 -60 118 -101 132 -92 6 4 10 18 8
+32 -3 22 3 28 48 44 28 11 59 28 70 40 l18 20 -107 -7 c-118 -8 -146 0 -166
+43 -17 37 -14 52 18 84 33 33 78 41 66 12 -14 -31 -16 -77 -5 -93 9 -13 14
+-11 36 14 21 24 25 37 21 68 -5 37 -4 38 36 49 56 15 192 6 242 -15 l40 -17
+-27 -20 c-66 -49 -1 -50 120 -2 80 31 92 40 92 64 0 33 -30 52 -72 45 -33 -5
+-51 1 -140 49 -226 121 -267 139 -327 139 -31 1 -68 -3 -83 -8 -24 -7 -32 -3
+-62 30 -62 67 -63 76 -30 145 53 111 38 151 -110 308 -88 93 -131 162 -142
+230 -11 67 0 84 99 164 142 113 222 232 261 384 46 178 -18 329 -188 441 -74
+48 -75 49 -51 60 35 16 31 28 -20 64 -31 22 -41 34 -32 40 21 14 168 59 252
+77 67 15 80 21 83 39 2 12 -33 92 -82 187 -100 193 -117 263 -35 144 28 -41
+102 -124 164 -185z m-430 -364 c-13 -21 19 -51 100 -97 78 -43 142 -93 119
+-93 -5 0 -34 7 -64 15 -69 19 -148 19 -180 0 -24 -14 -24 -14 21 -15 63 0 238
+-37 267 -56 24 -16 42 -52 42 -85 0 -17 -8 -14 -57 24 -92 71 -138 91 -213 90
+-36 0 -85 -8 -109 -17 -55 -20 -64 -14 -103 71 -37 82 -36 104 4 127 60 36
+190 63 173 36z m4328 -154 c76 -37 148 -103 127 -116 -27 -17 -107 -10 -174
+15 -91 34 -212 35 -310 1 -55 -19 -76 -22 -100 -14 -17 5 -37 12 -45 14 -17 6
+18 45 66 75 76 47 131 59 258 57 109 -3 125 -6 178 -32z m-4293 -282 c3 -28
+11 -58 17 -66 7 -8 12 -30 12 -49 -1 -40 10 -72 46 -128 33 -53 32 -67 -5 -80
+-78 -27 -118 37 -133 210 -10 105 -9 120 7 144 27 42 50 29 56 -31z m-53 -438
+c-4 -9 -11 -16 -17 -16 -11 0 -14 33 -3 44 11 10 26 -11 20 -28z m449 -922
+c60 -20 62 -23 44 -34 -23 -15 -104 -12 -126 5 -19 15 -19 15 0 30 24 18 22
+19 82 -1z m5843 -211 c82 -179 180 -472 218 -653 34 -160 38 -387 10 -517 -37
+-172 -107 -345 -191 -471 -81 -122 -215 -266 -232 -249 -2 2 10 37 27 78 188
+447 261 1077 185 1584 -15 98 -31 196 -35 217 -10 44 0 50 18 11z m-2689 -313
+c25 -333 24 -319 22 -342 -1 -10 -66 -56 -164 -115 -175 -106 -203 -121 -196
+-101 3 7 26 81 53 163 39 124 214 633 241 699 15 37 22 -12 44 -304z m110
+-832 c37 -172 37 -173 -56 -257 -80 -73 -143 -103 -230 -109 -105 -7 -132 9
+-183 108 -53 101 -53 132 -3 177 21 19 126 88 233 153 182 111 194 117 201 97
+3 -12 20 -88 38 -169z m-1046 -650 c67 -151 141 -236 331 -383 138 -106 239
+-173 309 -204 42 -18 47 -23 36 -36 -7 -9 -110 -81 -229 -160 -178 -118 -251
+-161 -416 -238 -110 -51 -250 -117 -312 -147 -62 -29 -118 -50 -126 -47 -7 3
+-24 34 -36 69 -28 77 -74 158 -252 445 -75 122 -140 231 -144 242 -6 19 -2 21
+42 21 63 0 171 24 230 50 35 15 99 72 238 209 182 180 271 261 285 261 4 0 24
+-37 44 -82z m-2387 -88 c-10 -81 -64 -284 -102 -378 -55 -136 -119 -238 -204
+-323 -108 -107 -162 -132 -306 -137 -109 -4 -111 -3 -175 30 -42 22 -76 48
+-98 78 l-35 45 114 7 c338 22 478 105 645 383 29 50 70 131 89 180 49 124 67
+165 73 165 3 0 2 -22 -1 -50z m1647 -1402 c-54 -78 -300 -358 -315 -358 -16 0
+-10 15 29 76 113 173 323 408 330 370 2 -10 -18 -49 -44 -88z"/>
+<path d="M3229 5845 c-108 -15 -150 -30 -198 -71 -49 -41 -111 -123 -111 -146
+0 -18 5 -20 40 -15 22 3 40 3 40 1 0 -2 -9 -26 -20 -53 -12 -32 -16 -52 -9
+-56 9 -6 7 -23 -8 -62 -3 -8 4 -13 20 -13 16 0 26 -7 30 -20 8 -30 33 -24 48
+12 15 37 122 148 142 148 12 0 11 -10 0 -56 -17 -66 -10 -118 17 -137 15 -11
+19 -21 14 -44 -5 -25 2 -39 44 -90 28 -33 66 -67 85 -76 39 -19 112 -22 152
+-7 39 15 108 74 140 121 27 38 28 41 13 72 -15 32 -15 34 8 54 13 12 24 33 24
+47 l0 25 38 -17 c58 -26 111 -62 122 -81 6 -13 3 -29 -11 -57 -98 -186 -404
+-264 -685 -174 -101 32 -143 37 -162 18 -21 -21 -13 -28 31 -28 49 0 82 -19
+91 -53 5 -20 11 -23 31 -19 14 2 31 0 38 -6 17 -14 145 -34 168 -27 11 4 26 1
+34 -5 9 -7 23 -9 37 -4 13 5 41 9 63 11 22 1 51 3 65 4 14 1 37 -1 52 -5 21
+-6 27 -4 33 13 5 18 14 21 55 21 43 0 49 3 52 23 3 19 10 22 51 25 49 3 53 7
+37 32 -11 17 4 30 38 30 15 0 22 6 22 19 0 14 16 27 55 45 40 18 56 31 61 51
+3 14 16 32 27 40 18 13 20 18 10 34 -11 17 -8 21 25 35 100 44 116 63 55 68
+-33 3 -38 6 -36 26 3 22 0 22 -45 16 -44 -6 -49 -4 -83 29 -58 56 -354 199
+-469 227 -116 28 -147 43 -70 36 30 -3 108 -23 173 -45 64 -22 117 -36 117
+-31 0 13 -23 24 -160 75 -142 53 -197 60 -331 40z"/>
+<path d="M3027 5303 c-3 -5 -2 -15 2 -22 7 -10 10 -10 16 -1 4 6 3 16 -3 22
+-5 5 -12 6 -15 1z"/>
+<path d="M2180 4462 c0 -11 136 -122 149 -122 19 0 12 46 -11 75 -12 15 -36
+34 -54 41 -37 15 -84 19 -84 6z"/>
+</g></g></svg>
@@ -0,0 +1,242 @@
+<svg version="1.2" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 500 500" width="500" height="500">
+	<style>
+		.s0 { fill: #f6f4f4 } 
+		.s1 { fill: #0b0303 } 
+		.s2 { fill: #ef0011 } 
+		.s3 { fill: #f3e2e2 } 
+		.s4 { fill: #f00212 } 
+		.s5 { fill: #ba000d } 
+		.s6 { fill: #faf1f1 } 
+		.s7 { fill: #0b0100 } 
+		.s8 { fill: #fbedee } 
+		.s9 { fill: #faeaea } 
+		.s10 { fill: #ab797d } 
+		.s11 { fill: #f8eaea } 
+		.s12 { fill: #902021 } 
+		.s13 { fill: #f9eeee } 
+		.s14 { fill: #f6ecec } 
+		.s15 { fill: #080201 } 
+		.s16 { fill: #150100 } 
+		.s17 { fill: #f2e7e7 } 
+		.s18 { fill: #fbe7e8 } 
+		.s19 { fill: #060101 } 
+		.s20 { fill: #f5e7e7 } 
+		.s21 { fill: #fa999e } 
+		.s22 { fill: #c46064 } 
+		.s23 { fill: #180300 } 
+		.s24 { fill: #f6dcdd } 
+		.s25 { fill: #f2e6e6 } 
+		.s26 { fill: #110200 } 
+		.s27 { fill: #eb0011 } 
+		.s28 { fill: #e20010 } 
+		.s29 { fill: #ea0011 } 
+		.s30 { fill: #760007 } 
+		.s31 { fill: #f00514 } 
+		.s32 { fill: #fcebeb } 
+		.s33 { fill: #ecd6d6 } 
+		.s34 { fill: #f5e3e3 } 
+		.s35 { fill: #f5e4e4 } 
+		.s36 { fill: #faf6f6 } 
+		.s37 { fill: #e50010 } 
+		.s38 { fill: #d5000f } 
+		.s39 { fill: #f2e2e3 } 
+		.s40 { fill: #ef1018 } 
+		.s41 { fill: #f4e8e9 } 
+		.s42 { fill: #ef0513 } 
+		.s43 { fill: #f5e5e5 } 
+		.s44 { fill: #f00413 } 
+		.s45 { fill: #f4e9ea } 
+		.s46 { fill: #ed0011 } 
+		.s47 { fill: #e80011 } 
+		.s48 { fill: #e60613 } 
+		.s49 { fill: #f0d6d6 } 
+		.s50 { fill: #fca9ac } 
+		.s51 { fill: #9c000c } 
+		.s52 { fill: #73393b } 
+	</style>
+	<g>
+		<path fill-rule="evenodd" class="s0" d="m166.5 52.5q3.5 0 7 0 2.75 2.99 1.5 7-21.27 45.61-20.5 96 39.99 2.76 72 26.5 7.87 6.86 13.5 15.5 42.88-56.39 103.5-92.5 47.35-25.46 101-25 14.52 0.38 23.5 11.5 3.19 7.74 2 16-1.81 7.18-4.5 14-1 0-1 1-5.04 6.05-9 13-1 0-1 1 0 0.5 0 1-12.42 12.15-28.5 19-6.02 36.27-41.5 45-0.83 2.75 0 5 19.02-12.85 41.5-9 10.85-8.09 23.5-13 15.01-6.37 31-2.5 14.09 7.43 14 23.5-2.83 23.25-15.5 43-6.42 9.92-14 19-10.04 8.8-19.5 18-72.02 48.88-156.5 27-19.63 9.6-41.5 10.5-4.59 1.27-9 3 2 1 4 2 20.09-1.11 35 12 25.46 6.95 37.5 30.5 1.26 5.69-1 11-3.38 3.79-7.5 6.5 5.74 10.07 1.5 20.5-7.55 7.47-17.5 3.5-11.01-5.34-22.5-9.5-18.26 10-38.5 13-15.5 0-31 0-26.62-4.54-51-17-4.17 1.33-8 3.5-7.23 5.87-15 11-8.62 2.58-13.5-4.5-1.82 2.32-4.5 3.5-6.06 2.24-12 3.5-7.5 0-15 0-27.42-2.56-50-18.5-18-17.25-23-41.5 0-11.5 0-23 4.12-22.7 25-33 6.95-16.67 22-26.5-20.39-20.8-14.5-49.5 7.01-26.98 28.5-44.5 7.56-5.27 15-10.5-13.09-30.88-7.5-64 3.16-15.57 14.5-26.5 6.85-2.48 8 4.5-6.59 39.53 11 75.5 7.99-0.49 16-2 2.42-34.57 14.5-67.5 8.51-22.23 27.5-36z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s1" d="m113.5 401.5q0.48-5.1-1-10-0.91 0.19-1 1-2.46 1.74-5 3.5 5.65 9.54-5 13-32.21 5.55-61-10-32.89-23.11-29.5-63.5 2.96-22.67 23.5-32 7.99-19.75 27-29.5-27.65-23.7-15.5-58.5 7.33-16.82 20.5-29.5 10.79-8.14 22-15.5-16.49-37.08-5.5-76 3.19-6.13 7.5-11.5 1.48-0.89 2 1-5.69 41.09 12.5 78.5 1 1 2 2 9.97-3.24 20.5-4 2 0 4 0 0-7.5 0-15 0.99-42.22 24.5-77 6.12-7.12 14-12-4.65 13.43-10 27-11.93 37.6-9.5 77 49.38 0.7 83.5 36 2.75 4.5 5.5 9 38.99-52.24 93-88.5 45.84-29.03 100-32.5 15.69-1.56 29 6.5 5.68 7.29 3.5 16.5-10.38 33.62-43.5 45-4.39 37.33-41 45-0.79 8.63-6 15.5 1.91 1.83 4.5 2.5 22.27-17.25 50.5-14.5 12.93-9.41 28-15 36.22-8.28 31.5 28.5-15.19 51.69-62.5 77.5-65.92 35.87-138 15.5-19.67 10.42-42 10.5-8.39 2.88-17 5 3.58 6.08 10 9 20.92-1.14 36 13 22.67 5.23 34.5 25.5 3.33 7.13-3.5 11.5-3.88 1.8-8 3 7.36 8.45 6.5 19.5-4.43 5.66-11.5 3.5-12.84-5.67-26-10.5-39.4 21.02-83 10.5-18.85-5.78-36.5-14.5-13.65 4.14-23.5 14.5-9.51 3.74-11-6.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s2" d="m153.5 173.5q24.62 1.46 46 13.5 12.11 8.1 17.5 21.5 0.74 2.45 0.5 5 0.09 0.81 1 1 1.48-4.9 1-10 5.04 10.48 1.5 22-9.81 27.86-35.5 42.5-26.17 14.97-56 19.5-2.77-0.4-2 1 2.86 1.27 6 1 25.64 1.53 48.5-10 0.34 10.08 2 20 1.08 5.76 5 10 1 1.5 0 3-31.11 20.84-68.5 17.5-23.7-5.7-32.5-28.5-4.39-9.18-3.5-19 15.41 6.23 32 4.5-20.68-6.39-39-18-34.81-27.22-12.5-65.5 11.84-14.83 29-23 4.21 7.66 11.5 12.5 3 1 6 0-26.04-34.62-29-78-0.13-8.46 2-16.5 1 6.5 2 13 3.43 39.53 24.5 73 2.03 2.28 4.5 4 0.5-1.25 1-2.5-1.27-6.54-5-12 0.5-0.75 1-1.5 9.72-3.43 20-4 0.55 10.34 8 17.5 1.94 0.74 4 0.5-17.8-64.6 16.5-122 0.98-1.79 1.5 0-28.21 56.64-13.5 118 1.08 1.43 2.5 0.5 2.21-4.98 2-10.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s3" d="m454.5 97.5q-18.37-2.97-37-1.5-16.14 2.08-32 5.5 32.38-14.09 67-7.5 1.98 1.22 2 3.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s4" d="m454.5 97.5q-1.33 11.18-8.5 20-21.81 26.28-55.5 32-1.11-0.2-2 0.5 2.31 2.82 5.5 4.5 1 2 0 4-9.56 11.3-19.5 20 19.71-8.72 31-27 2.68-0.43 5 1-14.24 30.97-48 36.5-9.93 1.71-20 1.5-6.8-0.48-13 1 5.81 6.92 14 11-10.78 16.03-27 26.5 27.16-7.4 38-33.5 4.34 1.35 9 1-9.08 23.84-33 33.5-18.45 6.41-38 7 22.59 8.92 45-1 12.05-5.52 24-11 9.01-1.79 17 2.5 5.28-4.38 11-8 12.8-6.07 27-5 0 0.5 0 1-19.34 2.69-34 15.5 0.5 0.25 1 0.5 17.79-8.09 36-15 2.71-0.79 5-2 2.5-1 5-2 5.53-4.04 11-8 11.7-4.18 24-6.5 7.78-1.36 15 1.5-2.97 18.45-13.5 34-34.92 49.37-94.5 62.5-59.27 12.45-108-23-15.53-12.52-21.5-31.5-2.47-14.26 4-27-3.15 24.41 14 42-4.92-10.28-7-22-1.97-17.63 7-33 47.28-69.5 125.5-100 15.86-3.42 32-5.5 18.63-1.47 37 1.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s5" d="m86.5 112.5q-1-6.5-2-13 0.7-5.34 3.5-10-1.8 11.32-1.5 23z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s6" d="m433.5 97.5q2.22-0.39 4 1-10 13.75-27 14-0.24-2.06 0.5-4 10.3-7.78 22.5-11z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s7" d="m407.5 101.5q2.55-0.24 5 0.5-52.87 18.31-84.5 64.5-6.94 7.95-17 11-9.38-2.38-5-11 40.38-48.62 101.5-65z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s8" d="m402.5 112.5q3 0 6 0-2.56 8.8-12 7-0.22-1.58 0.5-3 2.72-2.22 5.5-4z"/>
+	</g>
+	<g>
+	</g>
+	<g>
+	</g>
+	<g>
+	</g>
+	<g>
+	</g>
+	<g>
+	</g>
+	<g>
+	</g>
+	<g>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s9" d="m390.5 149.5q7.77 0.52 15 2-11.29 18.28-31 27 9.94-8.7 19.5-20 1-2 0-4-3.19-1.68-5.5-4.5 0.89-0.7 2-0.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s10" d="m131.5 145.5q0 7.5 0 15-2 0-4 0 1.06-1.36 3-1-0.48-7.29 1-14z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s11" d="m219.5 204.5q-1 4.5-2 9 0.24-2.55-0.5-5-5.39-13.4-17.5-21.5-21.38-12.04-46-13.5 0-2 0-4 36.7-0.86 61.5 26 3.06 4.11 4.5 9z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s12" d="m329.5 191.5q6.2-1.48 13-1-3.5 1-7 2-2.9-0.97-6-1z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s13" d="m329.5 191.5q3.1 0.03 6 1 9.55 1.31 19 3-10.84 26.1-38 33.5 16.22-10.47 27-26.5-8.19-4.08-14-11z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s14" d="m479.5 199.5q-7.22-2.86-15-1.5-12.3 2.32-24 6.5 15.6-13.11 36-11.5 3.63 2.26 3 6.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s15" d="m193.5 216.5q-12.01 1.52-22 8-2.83 1.29-5.5 3-4.79-4.57-6.5-11-5.04 2.2-9.5-1-3.47-6.4 3.5-3 4.4 0.05 8-2.5 9.22-9.73 21-16 6.3-3.24 12 1-2.9 1.22-6 1.5 2.61 5.74 4.5 12 0.75 3.97 0.5 8z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s16" d="m458.5 200.5q3.04-0.24 6 0.5-18.02 7.05-33 19-1 1-2 0 11.53-14.3 29-19.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s17" d="m178.5 202.5q6.85-0.63 4.5 6-7.6 5.09-6-4 1.08-0.82 1.5-2z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s18" d="m469.5 201.5q-2.26 13.65-14.5 22-0.47-2.11 1-4 7.08-8.82 13.5-18z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s19" d="m74.5 208.5q8.22-0.2 16 2.5 11.8 4.26 23.5 8.5 5.65-0.63 8-6 2.41 11.83-9.5 13 0.55 3.61 2 7-0.5 1-1 2-4.67-0.94-9.5-1-9.96 0.44-19.5 2.5-5.05-3.55-6.5-9.5-0.75-7.48-0.5-15-6.47 0.15-3-4z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s20" d="m429.5 212.5q-2.5 1-5 2-4 0-8 0-14.2-1.07-27 5 15.27-12.44 35-9.5 2.72 1.14 5 2.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s21" d="m219.5 204.5q0.48 5.1-1 10-0.91-0.19-1-1 1-4.5 2-9z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s22" d="m416.5 215.5q0-0.5 0-1 4 0 8 0-2.29 1.21-5 2-1.06-1.36-3-1z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s23" d="m416.5 215.5q1.94-0.36 3 1-18.21 6.91-36 15-0.5-0.25-1-0.5 14.66-12.81 34-15.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s24" d="m193.5 216.5q4.39 1.3 9 3-0.79 1.04-2 1.5-14.77-0.13-29 3.5 9.99-6.48 22-8z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s25" d="m98.5 219.5q6.09-0.98 6 5-3.04 0.24-6-0.5-1.84-2.24 0-4.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s26" d="m176.5 229.5q8.85-1.14 16 4-4.98 1.75-10 0-13.56 14.3-33 19.5-28.06 8.2-55 1 3.32-6.4 10-5.5-0.71 1.47-2 2.5 36.58 4.24 69-14 4.68-2.13 1-5 2.35-0.91 4-2.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s27" d="m231.5 238.5q1.31-0.2 2 1-3.13 28.62 15 51-16.25 6.75-27-7.5-1-1-2 0 14.73 29.34 46 18.5 1.79 0.52 0 1.5-37.63 16.82-50.5-22.5-5.1-26.48 16.5-42z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s28" d="m243.5 259.5q5.88 3.62 10.5 9 12.96 18.46 32.5 29.5-31.51-7.75-43-38.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s29" d="m203.5 266.5q1.31-0.2 2 1-2.48 22.08 12 39-6.99 1.35-14 0.5 4.59 4.08 10 7-8.71 0.28-14.5-6.5-16.98-22.76 4.5-41z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s27" d="m58.5 284.5q9.6-2.17 14.5 6 5.15 14.18-1 28-11.05-13.14-27.5-17.5 5.15-9.9 14-16.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s30" d="m129.5 288.5q2 1 4 2-3.14 0.27-6-1-0.77-1.4 2-1z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s31" d="m56.5 313.5q3.43 5.43 8 10-4.88 0.44-8 4-1.11-0.2-2 0.5 28.91 1.65 38 28.5 0.45 3.16-1 6-11.02-7.01-23-12.5-4.75-3.75-9.5-7.5 1.47 7.42 7 13 8.34 27.18 32 43 0.99 2.41-1.5 3.5-40.25 5.58-66.5-25.5-15.67-22.01-8-48 10.46-23.87 34.5-15z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s32" d="m45.5 317.5q4.03-0.25 8 0.5 2.46 4.16-2 6-6.04 2.01-9-3.5 1.26-1.85 3-3z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s33" d="m56.5 313.5q4.91 3.14 9.5 7 0.88 2.25-1.5 3-4.57-4.57-8-10z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s34" d="m198.5 319.5q-11.1 11.56-27 15.5-15.75 4.88-32 2.5 28.81-3.69 54-18.5 2.65-0.96 5 0.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s4" d="m198.5 319.5q1.44 0.68 2.5 2 2.41 8.23 6 16 1.2 2.64-0.5 5-30.65 21.41-68 18.5-25.16-6.17-32.5-30.5 6.96 4.99 15.5 6.5 8.99 0.75 18 0.5 16.25 2.38 32-2.5 15.9-3.94 27-15.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s35" d="m92.5 356.5q-9.09-26.85-38-28.5 0.89-0.7 2-0.5 25.47-4.89 35.5 19 0.75 4.98 0.5 10z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s36" d="m72.5 335.5q3.62-0.38 5 3-4.22 1.83-5-3z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s37" d="m223.5 336.5q5.59-0.48 11 1-4.04 4.16-8.5 8-5.99-3.8-2.5-9z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s38" d="m90.5 334.5q0.59-1.54 2-0.5 3.94 5.45 9 10 7 6 14 12-6.91-1.7-13-6-6.21-7.72-12-15.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s39" d="m261.5 346.5q-3.54-2.44-8-3.5-6.98-0.75-14-0.5 0.63-1.08 2-1.5 13.82-2.52 26 4-2.63 1.98-6 1.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s40" d="m239.5 342.5q7.02-0.25 14 0.5 4.46 1.06 8 3.5-5.2 2.35-10 5.5-3.88 4.65-9 7.5-9.89-3.09-9.5-13 2.36-3.63 6.5-4z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s41" d="m214.5 349.5q-21.43 15.48-48 16 22.82-5.9 43-18.5 3.64-1.12 5 2.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s42" d="m214.5 349.5q5.96 7.2 13.5 13 1 1 0 2-28.58 23.34-65.5 20.5-18.15-4.24-27.5-19.5 1.13 0.94 2.5 1.5 14.7 1.42 29-1.5 26.57-0.52 48-16z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s43" d="m302.5 373.5q-14.74-16.73-37-19-4.55 0.25-9 1 25.3-10.24 43.5 11 2.85 2.91 2.5 7z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s44" d="m302.5 373.5q0.21 2.44-2 3.5-28.69 7.6-50.5-12.5-0.06-6.71 6.5-9 4.45-0.75 9-1 22.26 2.27 37 19z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s45" d="m100.5 356.5q5.42 2.71 11 5.5-13.04 7.54-18.5 21.5-7.57-7.14-10.5-17 5.58 1.54 10 5.5 4.2 0.84 5.5-3.5 1.41-5.99 2.5-12z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s8" d="m83.5 394.5q-18.9-10.15-29.5-29-1.54-3.52-2-7 5.79 2.39 10 7 7.82 16.63 21.5 29z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s46" d="m232.5 365.5q17.6 6.19 10.5 23-10.6 10.42-25.5 11.5-25.94 3.21-49-9 36.75-1.65 64-25.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s47" d="m113.5 367.5q7.7-0.01 9.5 7-9.69 7.19-18.5 15.5-7.23 5.76-5.5-3.5 3.12-12.84 14.5-19z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s29" d="m126.5 380.5q7.88-0.4 12 6.5-8.5 7.25-17 14.5-5.62-12.55 5-21z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s48" d="m283.5 385.5q3.22 2.95 7 5.5 2.8 4.03 6 7.5 0.42 2.77-2 4-15.5-9.75-31-19.5-1.79-0.98 0-1.5 9.96 2.49 20 4z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s49" d="m283.5 385.5q8.71-1.27 11.5 7 1.22 2.9 1.5 6-3.2-3.47-6-7.5-3.78-2.55-7-5.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s50" d="m83.5 394.5q1.88-0.06 3 1.5-2.25 0.88-3-1.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s51" d="m258.5 392.5q3.51 0.41 0 2.5-2.33 1.93-5 2 2.61-2.28 5-4.5z"/>
+	</g>
+	<g>
+		<path fill-rule="evenodd" class="s52" d="m111.5 392.5q0.09-0.81 1-1 1.48 4.9 1 10-1-4.5-2-9z"/>
+	</g>
+</svg>
@@ -0,0 +1,7 @@
+<svg xmlns="http://www.w3.org/2000/svg" version="1.1" xmlns:xlink="http://www.w3.org/1999/xlink" width="512" height="512"><svg width="512" height="512" viewBox="0 0 512 512" fill="none" xmlns="http://www.w3.org/2000/svg">
+<rect width="512" height="512" fill="#131010"></rect>
+<path d="M320 224V352H192V224H320Z" fill="#5A5858"></path>
+<path fill-rule="evenodd" clip-rule="evenodd" d="M384 416H128V96H384V416ZM320 160H192V352H320V160Z" fill="white"></path>
+</svg><style>@media (prefers-color-scheme: light) { :root { filter: none; } }
+@media (prefers-color-scheme: dark) { :root { filter: none; } }
+</style></svg>
@@ -0,0 +1,9 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 800 800">
+  <rect width="800" height="800" rx="160" fill="#fff"/>
+  <path fill="#000" fill-rule="evenodd" d="
+    M165.29 165.29 H517.36 V400 H400 V517.36 H282.65 V634.72 H165.29 Z
+    M282.65 282.65 V400 H400 V282.65 Z
+  "/>
+  <path fill="#000" d="M517.36 400 H634.72 V634.72 H517.36 Z"/>
+</svg>
@@ -0,0 +1,9 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 800 800">
+  <rect width="800" height="800" rx="160" fill="#000"/>
+  <path fill="#fff" fill-rule="evenodd" d="
+    M165.29 165.29 H517.36 V400 H400 V517.36 H282.65 V634.72 H165.29 Z
+    M282.65 282.65 V400 H400 V282.65 Z
+  "/>
+  <path fill="#fff" d="M517.36 400 H634.72 V634.72 H517.36 Z"/>
+</svg>
@@ -406,6 +406,31 @@ export async function* pullModel(
  }
 }

+export interface ModelRecommendation {
+  model: string;
+  description: string;
+  context_length?: number;
+  max_output_tokens?: number;
+  vram_bytes?: number;
+}
+
+export interface ModelRecommendationsResponse {
+  recommendations: ModelRecommendation[];
+}
+
+export async function getModelRecommendations(): Promise<ModelRecommendation[]> {
+  const response = await fetch(
+    `${API_BASE}/api/experimental/model-recommendations`,
+  );
+  if (!response.ok) {
+    throw new Error(
+      `Failed to fetch model recommendations: ${response.statusText}`,
+    );
+  }
+  const data: ModelRecommendationsResponse = await response.json();
+  return data.recommendations || [];
+}
+
 export async function getInferenceCompute(): Promise<InferenceComputeResponse> {
  const response = await fetch(`${API_BASE}/api/v1/inference-compute`);
  if (!response.ok) {
@@ -480,13 +480,15 @@ function ChatForm({
      return;
    }

-    // Prepare attachments for submission
-    const attachmentsToSend: FileAttachment[] = message.attachments.map(
-      (att) => ({
+    // Prepare attachments for submission, excluding unsupported images
+    const attachmentsToSend: FileAttachment[] = message.attachments
+      .filter(
+        (att) => hasVisionCapability || !isImageFile(att.filename),
+      )
+      .map((att) => ({
        filename: att.filename,
        data: att.data || new Uint8Array(0), // Empty data for existing files
-      }),
-    );
+      }));

    const useWebSearch =
      supportsWebSearch && webSearchEnabled && !cloudDisabled;
@@ -736,10 +738,17 @@ function ChatForm({
        )}
        {(message.attachments.length > 0 || message.fileErrors.length > 0) && (
          <div className="flex gap-2 overflow-x-auto px-3 pt pb-3 w-full scrollbar-hide">
-            {message.attachments.map((attachment, index) => (
+            {message.attachments.map((attachment, index) => {
+              const isUnsupportedImage =
+                !hasVisionCapability && isImageFile(attachment.filename);
+              return (
              <div
                key={attachment.id}
-                className="group flex items-center gap-2 py-2 px-3 rounded-lg bg-neutral-50 dark:bg-neutral-700/50 hover:bg-neutral-100 dark:hover:bg-neutral-700 transition-colors flex-shrink-0"
+                className={`group flex items-center gap-2 py-2 px-3 rounded-lg transition-colors flex-shrink-0 ${
+                  isUnsupportedImage
+                    ? "bg-red-50 dark:bg-red-900/20 border border-red-200 dark:border-red-800"
+                    : "bg-neutral-50 dark:bg-neutral-700/50 hover:bg-neutral-100 dark:hover:bg-neutral-700"
+                }`}
              >
                {isImageFile(attachment.filename) ? (
                  <ImageThumbnail
@@ -764,9 +773,16 @@ function ChatForm({
                    />
                  </svg>
                )}
-                <span className="text-sm text-neutral-700 dark:text-neutral-300 max-w-[150px] truncate">
-                  {attachment.filename}
-                </span>
+                <div className="flex flex-col min-w-0">
+                  <span className={`text-sm max-w-36 truncate ${isUnsupportedImage ? "text-red-700 dark:text-red-300" : "text-neutral-700 dark:text-neutral-300"}`}>
+                    {attachment.filename}
+                  </span>
+                  {isUnsupportedImage && (
+                    <span className="text-xs text-red-600 dark:text-red-400 opacity-75">
+                      This model does not support images
+                    </span>
+                  )}
+                </div>
                <button
                  type="button"
                  onClick={() => removeFile(index)}
@@ -788,7 +804,8 @@ function ChatForm({
                  </svg>
                </button>
              </div>
-            ))}
+              );
+            })}
            {message.fileErrors.map((fileError, index) => (
              <div
                key={`error-${index}`}
@@ -6,12 +6,13 @@ import { getChat } from "@/api";
 import { Link } from "@/components/ui/link";
 import { useState, useRef, useEffect, useCallback, useMemo } from "react";
 import { ChatsResponse } from "@/gotypes";
-import { CogIcon } from "@heroicons/react/24/outline";
+import { CogIcon, RocketLaunchIcon } from "@heroicons/react/24/outline";

 // there's a hidden debug feature to copy a chat's data to the clipboard by
 // holding shift and clicking this many times within this many seconds
 const DEBUG_SHIFT_CLICKS_REQUIRED = 5;
 const DEBUG_SHIFT_CLICK_WINDOW_MS = 7000; // 7 seconds
+const launchSidebarRequestedKey = "ollama.launchSidebarRequested";

 interface ChatSidebarProps {
  currentChatId?: string;
@@ -267,9 +268,8 @@ export function ChatSidebar({ currentChatId }: ChatSidebarProps) {
        <Link
          href="/c/new"
          mask={{ to: "/" }}
-          className={`flex w-full items-center gap-3 rounded-lg px-2 py-2 text-left text-sm text-neutral-700 hover:bg-neutral-100 dark:hover:bg-neutral-800 dark:text-neutral-100 ${
-            currentChatId === "new" ? "bg-neutral-100 dark:bg-neutral-800" : ""
-          }`}
+          className={`flex w-full items-center gap-3 rounded-lg px-2 py-2 text-left text-sm text-neutral-700 hover:bg-neutral-100 dark:hover:bg-neutral-800 dark:text-neutral-100 ${currentChatId === "new" ? "bg-neutral-100 dark:bg-neutral-800" : ""
+            }`}
          draggable={false}
        >
          <svg
@@ -283,6 +283,23 @@ export function ChatSidebar({ currentChatId }: ChatSidebarProps) {
          </svg>
          <span className="truncate">New Chat</span>
        </Link>
+        <Link
+          to="/c/$chatId"
+          params={{ chatId: "launch" }}
+          onClick={() => {
+            if (currentChatId !== "launch") {
+              sessionStorage.setItem(launchSidebarRequestedKey, "1");
+            }
+          }}
+          className={`flex w-full items-center gap-3 rounded-lg px-2 py-2 text-left text-sm text-neutral-700 hover:bg-neutral-100 dark:hover:bg-neutral-800 dark:text-neutral-100 cursor-pointer ${currentChatId === "launch"
+            ? "bg-neutral-100 dark:bg-neutral-800"
+            : ""
+            }`}
+          draggable={false}
+        >
+          <RocketLaunchIcon className="h-5 w-5 stroke-current" />
+          <span className="truncate">Launch</span>
+        </Link>
        {isWindows && (
          <Link
            href="/settings"
@@ -304,19 +321,18 @@ export function ChatSidebar({ currentChatId }: ChatSidebarProps) {
              {group.chats.map((chat) => (
                <div
                  key={chat.id}
-                  className={`allow-context-menu flex items-center relative text-sm text-neutral-800 dark:text-neutral-400 rounded-lg hover:bg-neutral-100 dark:hover:bg-neutral-800 ${
-                    chat.id === currentChatId
-                      ? "bg-neutral-100 text-black dark:bg-neutral-800"
-                      : ""
-                  }`}
+                  className={`allow-context-menu flex items-center relative text-sm text-neutral-800 dark:text-neutral-400 rounded-lg hover:bg-neutral-100 dark:hover:bg-neutral-800 ${chat.id === currentChatId
+                    ? "bg-neutral-100 text-black dark:bg-neutral-800"
+                    : ""
+                    }`}
                  onMouseEnter={() => handleMouseEnter(chat.id)}
                  onContextMenu={(e) =>
                    handleContextMenu(
                      e,
                      chat.id,
                      chat.title ||
-                        chat.userExcerpt ||
-                        chat.createdAt.toLocaleString(),
+                      chat.userExcerpt ||
+                      chat.createdAt.toLocaleString(),
                    )
                  }
                >
@@ -10,6 +10,7 @@ interface CopyButtonProps {
  showLabels?: boolean;
  className?: string;
  title?: string;
+  onCopy?: () => void;
 }

 const CopyButton: React.FC<CopyButtonProps> = ({
@@ -20,6 +21,7 @@ const CopyButton: React.FC<CopyButtonProps> = ({
  showLabels = false,
  className = "",
  title = "",
+  onCopy,
 }) => {
  const [isCopied, setIsCopied] = useState(false);

@@ -48,12 +50,14 @@ const CopyButton: React.FC<CopyButtonProps> = ({
      }

      setIsCopied(true);
+      onCopy?.();
      setTimeout(() => setIsCopied(false), 2000);
    } catch (error) {
      console.error("Clipboard API failed, falling back to plain text", error);
      try {
        await navigator.clipboard.writeText(content);
        setIsCopied(true);
+        onCopy?.();
        setTimeout(() => setIsCopied(false), 2000);
      } catch (fallbackError) {
        console.error("Fallback copy also failed:", fallbackError);
@@ -0,0 +1,158 @@
+import { useSettings } from "@/hooks/useSettings";
+import CopyButton from "@/components/CopyButton";
+
+interface LaunchCommand {
+  id: string;
+  name: string;
+  command: string;
+  description: string;
+  icon: string;
+  darkIcon?: string;
+  iconClassName?: string;
+  borderless?: boolean;
+}
+
+const LAUNCH_COMMANDS: LaunchCommand[] = [
+  {
+    id: "claude",
+    name: "Claude Code",
+    command: "ollama launch claude",
+    description: "Anthropic's coding tool with subagents",
+    icon: "/launch-icons/claude-code.svg",
+    iconClassName: "h-7 w-7",
+  },
+  {
+    id: "chatgpt",
+    name: "ChatGPT",
+    command: "ollama launch chatgpt",
+    description: "Complete work with ChatGPT",
+    icon: "/launch-icons/codex-app.png",
+    iconClassName: "h-full w-full",
+  },
+  {
+    id: "hermes",
+    name: "Hermes Agent",
+    command: "ollama launch hermes",
+    description: "Self-improving AI agent built by Nous Research",
+    icon: "/launch-icons/hermes-agent.svg",
+    iconClassName: "h-7 w-7",
+  },
+  {
+    id: "openclaw",
+    name: "OpenClaw",
+    command: "ollama launch openclaw",
+    description: "Personal AI with 100+ skills",
+    icon: "/launch-icons/openclaw.svg",
+  },
+  {
+    id: "opencode",
+    name: "OpenCode",
+    command: "ollama launch opencode",
+    description: "Anomaly's open-source coding agent",
+    icon: "/launch-icons/opencode.svg",
+    iconClassName: "h-7 w-7 rounded",
+  },
+  {
+    id: "codex",
+    name: "Codex",
+    command: "ollama launch codex",
+    description: "OpenAI's open-source coding agent",
+    icon: "/launch-icons/codex.svg",
+    darkIcon: "/launch-icons/codex-dark.svg",
+    iconClassName: "h-7 w-7",
+  },
+  {
+    id: "copilot",
+    name: "Copilot CLI",
+    command: "ollama launch copilot",
+    description: "GitHub's AI coding agent for the terminal",
+    icon: "/launch-icons/copilot.svg",
+    darkIcon: "/launch-icons/copilot-dark.svg",
+    iconClassName: "h-7 w-7",
+  },
+  {
+    id: "droid",
+    name: "Droid",
+    command: "ollama launch droid",
+    description: "Factory's coding agent across terminal and IDEs",
+    icon: "/launch-icons/droid.svg",
+  },
+  {
+    id: "pi",
+    name: "Pi",
+    command: "ollama launch pi",
+    description: "Minimal AI agent toolkit with plugin support",
+    icon: "/launch-icons/pi.svg",
+    darkIcon: "/launch-icons/pi-dark.svg",
+    iconClassName: "h-7 w-7",
+  },
+];
+
+export default function LaunchCommands() {
+  const isWindows = navigator.platform.toLowerCase().includes("win");
+  const { setSettings } = useSettings();
+
+  const renderCommandCard = (item: LaunchCommand) => (
+    <div key={item.command} className="w-full text-left">
+      <div className="flex items-start gap-4 sm:gap-5">
+        <div
+          aria-hidden="true"
+          className={`flex h-10 w-10 shrink-0 items-center justify-center rounded-lg overflow-hidden ${item.borderless ? "" : "border border-neutral-200 bg-white dark:border-neutral-700 dark:bg-neutral-900"}`}
+        >
+          {item.darkIcon ? (
+            <picture>
+              <source srcSet={item.darkIcon} media="(prefers-color-scheme: dark)" />
+              <img src={item.icon} alt="" className={`${item.iconClassName ?? "h-8 w-8"} rounded-sm`} />
+            </picture>
+          ) : (
+            <img src={item.icon} alt="" className={item.borderless ? "h-full w-full rounded-xl" : `${item.iconClassName ?? "h-8 w-8"} rounded-sm`} />
+          )}
+        </div>
+
+        <div className="min-w-0 flex-1">
+          <span className="text-sm font-medium text-neutral-900 dark:text-neutral-100">
+            {item.name}
+          </span>
+          <p className="mt-0.5 text-xs text-neutral-500 dark:text-neutral-400">
+            {item.description}
+          </p>
+          <div className="mt-2 flex items-center gap-2 rounded-xl border-neutral-200 dark:border-neutral-700 bg-neutral-50 dark:bg-neutral-800 px-3 py-2">
+            <code className="min-w-0 flex-1 truncate text-xs text-neutral-600 dark:text-neutral-300">
+              {item.command}
+            </code>
+            <CopyButton
+              content={item.command}
+              size="md"
+              title="Copy command to clipboard"
+              className="text-neutral-500 dark:text-neutral-400 hover:text-neutral-700 dark:hover:text-neutral-200 hover:bg-neutral-200/60 dark:hover:bg-neutral-700/70"
+              onCopy={() => {
+                setSettings({ LastHomeView: item.id }).catch(() => { });
+              }}
+            />
+          </div>
+        </div>
+      </div>
+    </div>
+  );
+
+  return (
+    <main className="flex h-screen w-full flex-col relative">
+      <section
+        className={`flex-1 overflow-y-auto overscroll-contain relative min-h-0 ${isWindows ? "xl:pt-4" : "xl:pt-8"}`}
+      >
+        <div className="max-w-[730px] mx-auto w-full px-4 pt-4 pb-20 sm:px-6 sm:pt-6 sm:pb-24 lg:px-8 lg:pt-8 lg:pb-28">
+          <h1 className="text-xl font-semibold text-neutral-900 dark:text-neutral-100">
+            Launch
+          </h1>
+          <p className="mt-1 text-sm text-neutral-500 dark:text-neutral-400">
+            Copy a command and run it in your terminal.
+          </p>
+
+          <div className="mt-6 grid gap-7">
+            {LAUNCH_COMMANDS.map(renderCommandCard)}
+          </div>
+        </div>
+      </section>
+    </main>
+  );
+}
@@ -536,7 +536,7 @@ function ToolCallDisplay({
    let args: Record<string, unknown> | null = null;
    try {
      args = JSON.parse(toolCall.function.arguments) as Record<string, unknown>;
-    } catch (e) {
+    } catch {
      args = null;
    }
    const query = args && typeof args.query === "string" ? args.query : "";
@@ -562,7 +562,7 @@ function ToolCallDisplay({
    let args: Record<string, unknown> | null = null;
    try {
      args = JSON.parse(toolCall.function.arguments) as Record<string, unknown>;
-    } catch (e) {
+    } catch {
      args = null;
    }
    const url = args && typeof args.url === "string" ? args.url : "";
@@ -73,7 +73,7 @@ export default function MessageList({
                    ? String(args.url).trim()
                    : "";
              if (candidate) lastQuery = candidate;
-            } catch {}
+            } catch { /* ignored */ }
          }
        }
      }
@@ -273,6 +273,10 @@ export default function Settings() {
  }

  const isWindows = navigator.platform.toLowerCase().includes("win");
+  const handleCloseSettings = () => {
+    const chatId = settings.LastHomeView === "chat" ? "new" : "launch";
+    navigate({ to: "/c/$chatId", params: { chatId } });
+  };

  return (
    <main className="flex h-screen w-full flex-col select-none dark:bg-neutral-900">
@@ -286,7 +290,7 @@ export default function Settings() {
        >
          {isWindows && (
            <button
-              onClick={() => navigate({ to: "/" })}
+              onClick={handleCloseSettings}
              className="hover:bg-neutral-100 mr-3 dark:hover:bg-neutral-800 rounded-full p-1.5"
            >
              <ArrowLeftIcon className="w-5 h-5 dark:text-white" />
@@ -296,7 +300,7 @@ export default function Settings() {
        </h1>
        {!isWindows && (
          <button
-            onClick={() => navigate({ to: "/" })}
+            onClick={handleCloseSettings}
            className="p-1 hover:bg-neutral-100 mr-3 dark:hover:bg-neutral-800 rounded-full"
          >
            <XMarkIcon className="w-6 h-6 dark:text-white" />
@@ -0,0 +1,61 @@
+import { renderToStaticMarkup } from "react-dom/server";
+import type React from "react";
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+type MockStreamdownProps = {
+  children?: React.ReactNode;
+  components: {
+    img: React.ComponentType<React.ImgHTMLAttributes<HTMLImageElement>>;
+  };
+  rehypePlugins?: unknown[];
+};
+
+const streamdownMock = vi.hoisted(() =>
+  vi.fn((props: MockStreamdownProps) => props.children),
+);
+
+vi.mock("streamdown", () => ({
+  Streamdown: streamdownMock,
+  defaultRehypePlugins: {
+    katex: "katex",
+    raw: "raw",
+  },
+  defaultRemarkPlugins: {
+    gfm: "gfm",
+    math: "math",
+  },
+}));
+
+import StreamingMarkdownContent from "./StreamingMarkdownContent";
+
+describe("StreamingMarkdownContent", () => {
+  beforeEach(() => {
+    streamdownMock.mockClear();
+  });
+
+  it("does not enable raw HTML parsing", () => {
+    renderToStaticMarkup(
+      <StreamingMarkdownContent content="<iframe></iframe>" />,
+    );
+
+    const props = streamdownMock.mock.calls[0][0];
+    expect(props.rehypePlugins).toEqual(["katex"]);
+    expect(props.rehypePlugins).not.toContain("raw");
+  });
+
+  it("does not render markdown image src values", () => {
+    renderToStaticMarkup(
+      <StreamingMarkdownContent content="![secret](https://attacker.example/pixel?data=secret)" />,
+    );
+
+    const props = streamdownMock.mock.calls[0][0];
+    const Img = props.components.img;
+    const html = renderToStaticMarkup(
+      <Img alt="secret" src="https://attacker.example/pixel?data=secret" />,
+    );
+
+    expect(html).not.toContain("<img");
+    expect(html).not.toContain("attacker.example");
+    expect(html).toContain("secret");
+  });
+});
@@ -1,5 +1,9 @@
 import React from "react";
-import { Streamdown, defaultRemarkPlugins } from "streamdown";
+import {
+  Streamdown,
+  defaultRehypePlugins,
+  defaultRemarkPlugins,
+} from "streamdown";
 import remarkCitationParser from "@/utils/remarkCitationParser";
 import CopyButton from "./CopyButton";
 import type { BundledLanguage } from "shiki";
@@ -29,6 +33,8 @@ const extractText = (node: React.ReactNode): string => {
  return "";
 };

+const safeRehypePlugins = [defaultRehypePlugins.katex];
+
 const CodeBlock = React.memo(
  ({ children }: React.HTMLAttributes<HTMLPreElement>) => {
    // Extract code and language from children
@@ -210,9 +216,12 @@ const StreamingMarkdownContent: React.FC<StreamingMarkdownContentProps> =
          <Streamdown
            parseIncompleteMarkdown={isStreaming}
            isAnimating={isStreaming}
+            rehypePlugins={safeRehypePlugins}
            remarkPlugins={remarkPlugins}
            controls={false}
            components={{
+              img: ({ alt }: React.ImgHTMLAttributes<HTMLImageElement>) =>
+                alt ? <span>{alt}</span> : null,
              pre: CodeBlock,
              table: ({
                children,
@@ -65,7 +65,7 @@ export const BadgeButton = forwardRef(function BadgeButton(
    ),
  ref: React.ForwardedRef<HTMLElement>,
 ) {
-  let classes = clsx(
+  const classes = clsx(
    className,
    "group relative inline-flex rounded-md focus:not-data-focus:outline-hidden data-focus:outline-2 data-focus:outline-offset-2 data-focus:outline-blue-500",
  );
@@ -171,7 +171,7 @@ export const Button = forwardRef(function Button(
  { color, outline, plain, className, children, ...props }: ButtonProps,
  ref: React.ForwardedRef<HTMLElement>,
 ) {
-  let classes = clsx(
+  const classes = clsx(
    className,
    styles.base,
    outline
@@ -381,7 +381,7 @@ export const useSendMessage = (chatId: string) => {
                    role: "assistant",
                    content: "",
                    thinking: "",
-                    model: effectiveModel,
+                    model: effectiveModel.model,
                  }),
                );
                lastMessage = newMessages[newMessages.length - 1];
@@ -433,7 +433,7 @@ export const useSendMessage = (chatId: string) => {
                    role: "assistant",
                    content: "",
                    thinking: "",
-                    model: effectiveModel,
+                    model: effectiveModel.model,
                  }),
                );
                lastMessage = newMessages[newMessages.length - 1];
@@ -520,7 +520,7 @@ export const useSendMessage = (chatId: string) => {
                    thinkingTimeStart:
                      lastMessage.thinkingTimeStart || event.thinkingTimeStart,
                    thinkingTimeEnd: event.thinkingTimeEnd,
-                    model: selectedModel,
+                    model: selectedModel.model,
                  });
                  newMessages[newMessages.length - 1] = updatedMessage;
                } else {
@@ -533,7 +533,7 @@ export const useSendMessage = (chatId: string) => {
                      tool_calls: event.toolCalls,
                      thinkingTimeStart: event.thinkingTimeStart,
                      thinkingTimeEnd: event.thinkingTimeEnd,
-                      model: selectedModel,
+                      model: selectedModel.model,
                    }),
                  );
                }
@@ -699,7 +699,7 @@ export const useSendMessage = (chatId: string) => {
            queryClient.setQueryData(["chat", newId], {
              chat: new Chat({
                id: newId,
-                model: effectiveModel,
+                model: effectiveModel.model,
                messages: [
                  new Message({
                    role: "user",
@@ -0,0 +1,13 @@
+import { useQuery } from "@tanstack/react-query";
+import { getModelRecommendations } from "@/api";
+import type { ModelRecommendation } from "@/api";
+
+export function useFeaturedModels() {
+  return useQuery<ModelRecommendation[], Error>({
+    queryKey: ["modelRecommendations"],
+    queryFn: getModelRecommendations,
+    staleTime: 5 * 60 * 1000,
+    gcTime: 30 * 60 * 1000,
+    refetchOnWindowFocus: false,
+  });
+}
@@ -1,51 +1,49 @@
 import { useQuery } from "@tanstack/react-query";
 import { Model } from "@/gotypes";
 import { getModels } from "@/api";
-import { mergeModels } from "@/utils/mergeModels";
 import { useMemo } from "react";
 import { useCloudStatus } from "./useCloudStatus";
+import { useFeaturedModels } from "./useFeaturedModels";

 export function useModels(searchQuery = "") {
  const { cloudDisabled } = useCloudStatus();
+  const { data: recommendations, isLoading: recommendationsLoading } =
+    useFeaturedModels();
  const localQuery = useQuery<Model[], Error>({
    queryKey: ["models", searchQuery],
    queryFn: () => getModels(searchQuery),
-    gcTime: 10 * 60 * 1000, // Keep in cache for 10 minutes
+    gcTime: 10 * 60 * 1000,
    retry: 10,
-    // exponential backoff, starting at 100ms and capping at 5s
    retryDelay: (attemptIndex) => Math.min(100 * 2 ** attemptIndex, 5000),
    refetchOnWindowFocus: true,
-    refetchInterval: 30 * 1000, // Refetch every 30 seconds to keep models updated
+    refetchInterval: 30 * 1000,
    refetchIntervalInBackground: true,
  });

  const allModels = useMemo(() => {
-    const models = mergeModels(localQuery.data || [], cloudDisabled);
+    const local = localQuery.data || [];
+    const featured = (recommendations || []).map((r) => r.model);
+    const featuredSet = new Set(featured);

-    if (searchQuery && searchQuery.trim()) {
-      const query = searchQuery.toLowerCase().trim();
-      const filteredModels = models.filter((model) =>
-        model.model.toLowerCase().includes(query),
-      );
+    // Recommended models first (using the local copy when downloaded),
+    // then everything else from /api/tags in tags order.
+    const recommended = featured.map(
+      (name) =>
+        local.find((m) => m.model === name) || new Model({ model: name }),
+    );
+    const rest = local.filter((m) => !featuredSet.has(m.model));
+    const merged = [...recommended, ...rest];

-      const seen = new Set<string>();
-      return filteredModels.filter((model) => {
-        const currentModel = model.model.toLowerCase();
-        if (seen.has(currentModel)) {
-          return false;
-        }
-        seen.add(currentModel);
-        return true;
-      });
-    }
-
-    return models;
-  }, [localQuery.data, searchQuery, cloudDisabled]);
+    const visible = cloudDisabled
+      ? merged.filter((m) => !m.isCloud())
+      : merged;
+    return filterBySearch(visible, searchQuery);
+  }, [localQuery.data, searchQuery, cloudDisabled, recommendations]);

  return {
    ...localQuery,
    data: allModels,
-    isLoading: localQuery.isLoading,
+    isLoading: localQuery.isLoading || recommendationsLoading,
  };
 }

@@ -53,3 +51,16 @@ export function useRefetchModels() {
  const { refetch } = useModels();
  return refetch;
 }
+
+function filterBySearch(models: Model[], query: string): Model[] {
+  const q = query.trim().toLowerCase();
+  if (!q) return models;
+
+  const seen = new Set<string>();
+  return models.filter((m) => {
+    const name = m.model.toLowerCase();
+    if (!name.includes(q) || seen.has(name)) return false;
+    seen.add(name);
+    return true;
+  });
+}
@@ -4,7 +4,6 @@ import { useModels } from "./useModels";
 import { useChat } from "./useChats";
 import { useSettings } from "./useSettings.ts";
 import { Model } from "@/gotypes";
-import { FEATURED_MODELS } from "@/utils/mergeModels";
 import { getTotalVRAM } from "@/utils/vram.ts";
 import { getInferenceCompute } from "@/api";
 import { useCloudStatus } from "./useCloudStatus";
@@ -92,9 +91,7 @@ export function useSelectedModel(currentChatId?: string, searchQuery?: string) {
      (settings.selectedModel &&
        new Model({
          model: settings.selectedModel,
-          cloud: FEATURED_MODELS.some(
-            (f) => f.endsWith("cloud") && f === settings.selectedModel,
-          ),
+          cloud: settings.selectedModel.endsWith("cloud"),
          ollama_host: false,
        })) ||
      null
@@ -9,6 +9,7 @@ interface SettingsState {
  webSearchEnabled: boolean;
  selectedModel: string;
  sidebarOpen: boolean;
+  lastHomeView: string;
  thinkEnabled: boolean;
  thinkLevel: string;
 }
@@ -21,6 +22,7 @@ type SettingsUpdate = Partial<{
  ThinkLevel: string;
  SelectedModel: string;
  SidebarOpen: boolean;
+  LastHomeView: string;
 }>;

 export function useSettings() {
@@ -50,6 +52,7 @@ export function useSettings() {
      thinkLevel: settingsData?.settings?.ThinkLevel ?? "none",
      selectedModel: settingsData?.settings?.SelectedModel ?? "",
      sidebarOpen: settingsData?.settings?.SidebarOpen ?? false,
+      lastHomeView: settingsData?.settings?.LastHomeView ?? "launch",
    }),
    [settingsData?.settings],
  );
@@ -4,12 +4,37 @@ import Chat from "@/components/Chat";
 import { getChat } from "@/api";
 import { SidebarLayout } from "@/components/layout/layout";
 import { ChatSidebar } from "@/components/ChatSidebar";
+import LaunchCommands from "@/components/LaunchCommands";
+import { useEffect, useRef } from "react";
+import { useSettings } from "@/hooks/useSettings";
+
+const launchSidebarRequestedKey = "ollama.launchSidebarRequested";
+const launchSidebarSeenKey = "ollama.launchSidebarSeen";
+const fallbackSessionState = new Map<string, string>();
+
+function getSessionState() {
+  if (typeof sessionStorage !== "undefined") {
+    return sessionStorage;
+  }
+
+  return {
+    getItem(key: string) {
+      return fallbackSessionState.get(key) ?? null;
+    },
+    setItem(key: string, value: string) {
+      fallbackSessionState.set(key, value);
+    },
+    removeItem(key: string) {
+      fallbackSessionState.delete(key);
+    },
+  };
+}

 export const Route = createFileRoute("/c/$chatId")({
  component: RouteComponent,
  loader: async ({ context, params }) => {
-    // Skip loading for "new" chat
-    if (params.chatId !== "new") {
+    // Skip loading for special non-chat views
+    if (params.chatId !== "new" && params.chatId !== "launch") {
      context.queryClient.ensureQueryData({
        queryKey: ["chat", params.chatId],
        queryFn: () => getChat(params.chatId),
@@ -21,13 +46,70 @@ export const Route = createFileRoute("/c/$chatId")({

 function RouteComponent() {
  const { chatId } = Route.useParams();
+  const { settingsData, setSettings } = useSettings();
+  const previousChatIdRef = useRef<string | null>(null);

-  // Always call hooks at the top level - use a flag to skip data when chatId is "new"
+  // Always call hooks at the top level - use a flag to skip data when chatId is a special view
  const {
    data: chatData,
    isLoading: chatLoading,
    error: chatError,
-  } = useChat(chatId === "new" ? "" : chatId);
+  } = useChat(chatId === "new" || chatId === "launch" ? "" : chatId);
+
+  useEffect(() => {
+    if (!settingsData) {
+      return;
+    }
+
+    const previousChatId = previousChatIdRef.current;
+    previousChatIdRef.current = chatId;
+
+    if (chatId === "launch") {
+      const sessionState = getSessionState();
+      const shouldOpenSidebar =
+        previousChatId !== "launch" &&
+        (() => {
+          if (sessionState.getItem(launchSidebarRequestedKey) === "1") {
+            sessionState.removeItem(launchSidebarRequestedKey);
+            sessionState.setItem(launchSidebarSeenKey, "1");
+            return true;
+          }
+
+          if (sessionState.getItem(launchSidebarSeenKey) !== "1") {
+            sessionState.setItem(launchSidebarSeenKey, "1");
+            return true;
+          }
+
+          return false;
+        })();
+      const updates: { LastHomeView?: string; SidebarOpen?: boolean } = {};
+
+      if (settingsData.LastHomeView !== "launch") {
+        updates.LastHomeView = "launch";
+      }
+
+      if (shouldOpenSidebar && !settingsData.SidebarOpen) {
+        updates.SidebarOpen = true;
+      }
+
+      if (Object.keys(updates).length === 0) {
+        return;
+      }
+
+      setSettings(updates).catch(() => {
+        // Best effort persistence for home view preference.
+      });
+      return;
+    }
+
+    if (settingsData.LastHomeView === "chat") {
+      return;
+    }
+
+    setSettings({ LastHomeView: "chat" }).catch(() => {
+      // Best effort persistence for home view preference.
+    });
+  }, [chatId, settingsData, setSettings]);

  // Handle "new" chat case - just use Chat component which handles everything
  if (chatId === "new") {
@@ -38,6 +120,14 @@ function RouteComponent() {
    );
  }

+  if (chatId === "launch") {
+    return (
+      <SidebarLayout sidebar={<ChatSidebar currentChatId={chatId} />}>
+        <LaunchCommands />
+      </SidebarLayout>
+    );
+  }
+
  // Handle existing chat case
  if (chatLoading) {
    return (
@@ -1,10 +1,18 @@
 import { createFileRoute, redirect } from "@tanstack/react-router";
+import { getSettings } from "@/api";

 export const Route = createFileRoute("/")({
-  beforeLoad: () => {
+  beforeLoad: async ({ context }) => {
+    const settingsData = await context.queryClient.ensureQueryData({
+      queryKey: ["settings"],
+      queryFn: getSettings,
+    });
+    const chatId =
+      settingsData?.settings?.LastHomeView === "chat" ? "new" : "launch";
+
    throw redirect({
      to: "/c/$chatId",
-      params: { chatId: "new" },
+      params: { chatId },
      mask: {
        to: "/",
      },
@@ -0,0 +1,57 @@
+import { describe, expect, it, vi, beforeEach } from "vitest";
+import { copyTextToClipboard } from "./clipboard";
+
+describe("copyTextToClipboard", () => {
+  beforeEach(() => {
+    vi.restoreAllMocks();
+  });
+
+  it("copies via Clipboard API when available", async () => {
+    const writeText = vi.fn().mockResolvedValue(undefined);
+    vi.stubGlobal("navigator", {
+      clipboard: {
+        writeText,
+      },
+    });
+
+    const copied = await copyTextToClipboard("ollama launch claude");
+
+    expect(copied).toBe(true);
+    expect(writeText).toHaveBeenCalledWith("ollama launch claude");
+  });
+
+  it("falls back to execCommand when Clipboard API fails", async () => {
+    const writeText = vi.fn().mockRejectedValue(new Error("not allowed"));
+    vi.stubGlobal("navigator", {
+      clipboard: {
+        writeText,
+      },
+    });
+
+    const textarea = {
+      value: "",
+      setAttribute: vi.fn(),
+      style: {} as Record<string, string>,
+      focus: vi.fn(),
+      select: vi.fn(),
+    };
+    const appendChild = vi.fn();
+    const removeChild = vi.fn();
+    const execCommand = vi.fn().mockReturnValue(true);
+    vi.stubGlobal("document", {
+      createElement: vi.fn().mockReturnValue(textarea),
+      body: {
+        appendChild,
+        removeChild,
+      },
+      execCommand,
+    });
+
+    const copied = await copyTextToClipboard("ollama launch openclaw");
+
+    expect(copied).toBe(true);
+    expect(execCommand).toHaveBeenCalledWith("copy");
+    expect(appendChild).toHaveBeenCalled();
+    expect(removeChild).toHaveBeenCalled();
+  });
+});
@@ -0,0 +1,30 @@
+export async function copyTextToClipboard(text: string): Promise<boolean> {
+  try {
+    await navigator.clipboard.writeText(text);
+    return true;
+  } catch (clipboardError) {
+    console.error(
+      "Clipboard API failed, falling back to execCommand",
+      clipboardError,
+    );
+  }
+
+  try {
+    const textarea = document.createElement("textarea");
+    textarea.value = text;
+    textarea.setAttribute("readonly", "true");
+    textarea.style.position = "fixed";
+    textarea.style.left = "-9999px";
+    textarea.style.opacity = "0";
+    document.body.appendChild(textarea);
+    textarea.focus();
+    textarea.select();
+
+    const copied = document.execCommand("copy");
+    document.body.removeChild(textarea);
+    return copied;
+  } catch (fallbackError) {
+    console.error("Fallback copy failed", fallbackError);
+    return false;
+  }
+}
@@ -29,13 +29,15 @@ describe("fileValidation", () => {
      expect(result.valid).toBe(true);
    });

-    it("should reject WebP images when vision capability is disabled", () => {
+    it("should accept images regardless of vision capability", () => {
+      // Vision capability check is handled at the UI layer (ChatForm),
+      // not at validation time, so users can switch models without
+      // needing to re-upload files.
      const file = createMockFile("test.webp", 1024, "image/webp");
      const result = validateFile(file, {
        hasVisionCapability: false,
      });
-      expect(result.valid).toBe(false);
-      expect(result.error).toBe("This model does not support images");
+      expect(result.valid).toBe(true);
    });

    it("should accept PNG images when vision capability is enabled", () => {
@@ -63,7 +63,6 @@ export function validateFile(
  const {
    maxFileSize = 10,
    allowedExtensions = [...TEXT_FILE_EXTENSIONS, ...IMAGE_EXTENSIONS],
-    hasVisionCapability = false,
    customValidator,
  } = options;

@@ -83,10 +82,6 @@ export function validateFile(
    return { valid: false, error: "File type not supported" };
  }

-  if (IMAGE_EXTENSIONS.includes(fileExtension) && !hasVisionCapability) {
-    return { valid: false, error: "This model does not support images" };
-  }
-
  // File size validation
  if (file.size > MAX_FILE_SIZE) {
    return { valid: false, error: "File too large" };
@@ -1,128 +0,0 @@
-import { describe, it, expect } from "vitest";
-import { Model } from "@/gotypes";
-import { mergeModels, FEATURED_MODELS } from "@/utils/mergeModels";
-import "@/api";
-
-describe("Model merging logic", () => {
-  it("should handle cloud models with -cloud suffix", () => {
-    const localModels: Model[] = [
-      new Model({ model: "gpt-oss:120b-cloud" }),
-      new Model({ model: "llama3:latest" }),
-      new Model({ model: "mistral:latest" }),
-    ];
-
-    const merged = mergeModels(localModels);
-
-    // First verify cloud models are first and in FEATURED_MODELS order
-    const cloudModels = FEATURED_MODELS.filter((m: string) =>
-      m.endsWith("cloud"),
-    );
-    for (let i = 0; i < cloudModels.length; i++) {
-      expect(merged[i].model).toBe(cloudModels[i]);
-      expect(merged[i].isCloud()).toBe(true);
-    }
-
-    // Then verify non-cloud featured models are next and in FEATURED_MODELS order
-    const nonCloudFeatured = FEATURED_MODELS.filter(
-      (m: string) => !m.endsWith("cloud"),
-    );
-    for (let i = 0; i < nonCloudFeatured.length; i++) {
-      const model = merged[i + cloudModels.length];
-      expect(model.model).toBe(nonCloudFeatured[i]);
-      expect(model.isCloud()).toBe(false);
-    }
-
-    // Verify local models are preserved and come after featured models
-    const featuredCount = FEATURED_MODELS.length;
-    expect(merged[featuredCount].model).toBe("llama3:latest");
-    expect(merged[featuredCount + 1].model).toBe("mistral:latest");
-
-    // Length should be exactly featured models plus our local models
-    expect(merged.length).toBe(FEATURED_MODELS.length + 2);
-  });
-
-  it("should hide cloud models when cloud is disabled", () => {
-    const localModels: Model[] = [
-      new Model({ model: "gpt-oss:120b-cloud" }),
-      new Model({ model: "llama3:latest" }),
-      new Model({ model: "mistral:latest" }),
-    ];
-
-    const merged = mergeModels(localModels, true); // cloud disabled = true
-
-    // No cloud models should be present
-    const cloudModels = merged.filter((m) => m.isCloud());
-    expect(cloudModels.length).toBe(0);
-
-    // Should have non-cloud featured models
-    const nonCloudFeatured = FEATURED_MODELS.filter(
-      (m) => !m.endsWith("cloud"),
-    );
-    for (let i = 0; i < nonCloudFeatured.length; i++) {
-      const model = merged[i];
-      expect(model.model).toBe(nonCloudFeatured[i]);
-      expect(model.isCloud()).toBe(false);
-    }
-
-    // Local models should be preserved
-    const featuredCount = nonCloudFeatured.length;
-    expect(merged[featuredCount].model).toBe("llama3:latest");
-    expect(merged[featuredCount + 1].model).toBe("mistral:latest");
-  });
-
-  it("should handle empty input", () => {
-    const merged = mergeModels([]);
-
-    // First verify cloud models are first and in FEATURED_MODELS order
-    const cloudModels = FEATURED_MODELS.filter((m) => m.endsWith("cloud"));
-    for (let i = 0; i < cloudModels.length; i++) {
-      expect(merged[i].model).toBe(cloudModels[i]);
-      expect(merged[i].isCloud()).toBe(true);
-    }
-
-    // Then verify non-cloud featured models are next and in FEATURED_MODELS order
-    const nonCloudFeatured = FEATURED_MODELS.filter(
-      (m) => !m.endsWith("cloud"),
-    );
-    for (let i = 0; i < nonCloudFeatured.length; i++) {
-      const model = merged[i + cloudModels.length];
-      expect(model.model).toBe(nonCloudFeatured[i]);
-      expect(model.isCloud()).toBe(false);
-    }
-
-    // Length should be exactly FEATURED_MODELS length
-    expect(merged.length).toBe(FEATURED_MODELS.length);
-  });
-
-  it("should sort models correctly", () => {
-    const localModels: Model[] = [
-      new Model({ model: "zephyr:latest" }),
-      new Model({ model: "alpha:latest" }),
-      new Model({ model: "gpt-oss:120b-cloud" }),
-    ];
-
-    const merged = mergeModels(localModels);
-
-    // First verify cloud models are first and in FEATURED_MODELS order
-    const cloudModels = FEATURED_MODELS.filter((m) => m.endsWith("cloud"));
-    for (let i = 0; i < cloudModels.length; i++) {
-      expect(merged[i].model).toBe(cloudModels[i]);
-      expect(merged[i].isCloud()).toBe(true);
-    }
-
-    // Then verify non-cloud featured models are next and in FEATURED_MODELS order
-    const nonCloudFeatured = FEATURED_MODELS.filter(
-      (m) => !m.endsWith("cloud"),
-    );
-    for (let i = 0; i < nonCloudFeatured.length; i++) {
-      const model = merged[i + cloudModels.length];
-      expect(model.model).toBe(nonCloudFeatured[i]);
-      expect(model.isCloud()).toBe(false);
-    }
-
-    // Non-featured local models should be at the end in alphabetical order
-    const featuredCount = FEATURED_MODELS.length;
-    expect(merged[featuredCount].model).toBe("alpha:latest");
-    expect(merged[featuredCount + 1].model).toBe("zephyr:latest");
-  });
-});
@@ -1,101 +0,0 @@
-import { Model } from "@/gotypes";
-
-// Featured models list (in priority order)
-export const FEATURED_MODELS = [
-  "gpt-oss:120b-cloud",
-  "gpt-oss:20b-cloud",
-  "deepseek-v3.1:671b-cloud",
-  "qwen3-coder:480b-cloud",
-  "qwen3-vl:235b-cloud",
-  "minimax-m2:cloud",
-  "glm-4.6:cloud",
-  "gpt-oss:120b",
-  "gpt-oss:20b",
-  "gemma3:27b",
-  "gemma3:12b",
-  "gemma3:4b",
-  "gemma3:1b",
-  "deepseek-r1:8b",
-  "qwen3-coder:30b",
-  "qwen3-vl:30b",
-  "qwen3-vl:8b",
-  "qwen3-vl:4b",
-  "qwen3:30b",
-  "qwen3:8b",
-  "qwen3:4b",
-];
-
-function alphabeticalSort(a: Model, b: Model): number {
-  return a.model.toLowerCase().localeCompare(b.model.toLowerCase());
-}
-
-//Merges models, sorting cloud models first, then other models
-export function mergeModels(
-  localModels: Model[],
-  hideCloudModels: boolean = false,
-): Model[] {
-  const allModels = (localModels || []).map((model) => model);
-
-  // 1. Get cloud models from local models and featured list
-  const cloudModels = [...allModels.filter((m) => m.isCloud())];
-
-  // Add any cloud models from FEATURED_MODELS that aren't in local models
-  FEATURED_MODELS.filter((f) => f.endsWith("cloud")).forEach((cloudModel) => {
-    if (!cloudModels.some((m) => m.model === cloudModel)) {
-      cloudModels.push(new Model({ model: cloudModel }));
-    }
-  });
-
-  // 2. Get other featured models (non-cloud)
-  const featuredModels = FEATURED_MODELS.filter(
-    (f) => !f.endsWith("cloud"),
-  ).map((model) => {
-    // Check if this model exists in local models
-    const localMatch = allModels.find(
-      (m) => m.model.toLowerCase() === model.toLowerCase(),
-    );
-
-    if (localMatch) return localMatch;
-
-    return new Model({
-      model,
-    });
-  });
-
-  // 3. Get remaining local models that aren't featured and aren't cloud models
-  const remainingModels = allModels.filter(
-    (model) =>
-      !model.isCloud() &&
-      !FEATURED_MODELS.some(
-        (f) => f.toLowerCase() === model.model.toLowerCase(),
-      ),
-  );
-
-  cloudModels.sort((a, b) => {
-    const aIndex = FEATURED_MODELS.indexOf(a.model);
-    const bIndex = FEATURED_MODELS.indexOf(b.model);
-
-    // If both are featured, sort by their position in FEATURED_MODELS
-    if (aIndex !== -1 && bIndex !== -1) {
-      return aIndex - bIndex;
-    }
-
-    // If only one is featured, featured model comes first
-    if (aIndex !== -1 && bIndex === -1) return -1;
-    if (aIndex === -1 && bIndex !== -1) return 1;
-
-    // If neither is featured, sort alphabetically
-    return a.model.toLowerCase().localeCompare(b.model.toLowerCase());
-  });
-
-  featuredModels.sort(
-    (a, b) =>
-      FEATURED_MODELS.indexOf(a.model) - FEATURED_MODELS.indexOf(b.model),
-  );
-
-  remainingModels.sort(alphabeticalSort);
-
-  return hideCloudModels
-    ? [...featuredModels, ...remainingModels]
-    : [...cloudModels, ...featuredModels, ...remainingModels];
-}
@@ -302,6 +302,7 @@ func (s *Server) Handler() http.Handler {
 	mux.Handle("HEAD /api/version", ollamaProxy)
 	mux.Handle("POST /api/me", ollamaProxy)
 	mux.Handle("POST /api/signout", ollamaProxy)
+	mux.Handle("GET /api/experimental/model-recommendations", ollamaProxy)

 	// React app - catch all non-API routes and serve the React app
 	mux.Handle("GET /", s.appHandler())
@@ -342,8 +343,18 @@ func (t *userAgentTransport) RoundTrip(req *http.Request) (*http.Response, error

 // httpClient returns an HTTP client that automatically adds the User-Agent header
 func (s *Server) httpClient() *http.Client {
+	return userAgentHTTPClient(10 * time.Second)
+}
+
+// inferenceClient uses almost the same HTTP client, but without a timeout so
+// long requests aren't truncated
+func (s *Server) inferenceClient() *api.Client {
+	return api.NewClient(envconfig.Host(), userAgentHTTPClient(0))
+}
+
+func userAgentHTTPClient(timeout time.Duration) *http.Client {
 	return &http.Client{
-		Timeout: 10 * time.Second,
+		Timeout: timeout,
 		Transport: &userAgentTransport{
 			base: http.DefaultTransport,
 		},
@@ -563,6 +574,18 @@ func (s *Server) getError(err error) responses.ErrorEvent {
 	}
 }

+func userMessageText(messages []store.Message) string {
+	var b strings.Builder
+	for _, message := range messages {
+		if message.Role != "user" {
+			continue
+		}
+		b.WriteString(message.Content)
+		b.WriteByte('\n')
+	}
+	return b.String()
+}
+
 func (s *Server) browserState(chat *store.Chat) (*responses.BrowserStateData, bool) {
 	if len(chat.BrowserState) > 0 {
 		var st responses.BrowserStateData
@@ -721,11 +744,7 @@ func (s *Server) chat(w http.ResponseWriter, r *http.Request) error {
 	_, cancelLoading := context.WithCancel(ctx)
 	loading := false

-	c, err := api.ClientFromEnvironment()
-	if err != nil {
-		cancelLoading()
-		return err
-	}
+	c := s.inferenceClient()

 	// Check if the model exists locally by trying to show it
 	// TODO (jmorganca): skip this round trip and instead just act
@@ -832,6 +851,7 @@ func (s *Server) chat(w http.ResponseWriter, r *http.Request) error {
 	// Note: Skip agent/tools mode if user has attachments, as the agent doesn't handle file attachments properly
 	registry := tools.NewRegistry()
 	var browser *tools.Browser
+	ctx = tools.WithAllowedDirectURLs(ctx, userMessageText(chat.Messages))

 	if !hasAttachments {
 		WebSearchEnabled := req.WebSearch != nil && *req.WebSearch
@@ -1682,7 +1702,6 @@ func supportsBrowserTools(model string) bool {
 	return strings.HasPrefix(strings.ToLower(model), "gpt-oss")
 }

-
 // buildChatRequest converts store.Chat to api.ChatRequest
 func (s *Server) buildChatRequest(chat *store.Chat, model string, think any, availableTools []map[string]any) (*api.ChatRequest, error) {
 	var msgs []api.Message
@@ -15,6 +15,7 @@ import (
 	"sync/atomic"
 	"testing"

+	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/app/store"
 	"github.com/ollama/ollama/app/updater"
 )
@@ -526,6 +527,33 @@ func TestUserAgentTransport(t *testing.T) {
 	t.Logf("User-Agent transport successfully set: %s", receivedUA)
 }

+func TestInferenceClientUsesUserAgent(t *testing.T) {
+	var gotUserAgent atomic.Value
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotUserAgent.Store(r.Header.Get("User-Agent"))
+		w.Header().Set("Content-Type", "application/json")
+		w.Write([]byte(`{}`))
+	}))
+	defer ts.Close()
+
+	t.Setenv("OLLAMA_HOST", ts.URL)
+
+	server := &Server{}
+	client := server.inferenceClient()
+
+	_, err := client.Show(context.Background(), &api.ShowRequest{Model: "test"})
+	if err != nil {
+		t.Fatalf("show request failed: %v", err)
+	}
+
+	receivedUA, _ := gotUserAgent.Load().(string)
+	expectedUA := userAgent()
+
+	if receivedUA != expectedUA {
+		t.Errorf("User-Agent mismatch\nExpected: %s\nReceived: %s", expectedUA, receivedUA)
+	}
+}
+
 func TestSupportsBrowserTools(t *testing.T) {
 	tests := []struct {
 		model string
@@ -5,6 +5,8 @@ package updater
 import (
 	"context"
 	"crypto/rand"
+	"crypto/sha256"
+	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -169,22 +171,20 @@ func (u *Updater) DownloadNewRelease(ctx context.Context, updateResp UpdateRespo
 	if err != nil {
 		return fmt.Errorf("error checking update: %w", err)
 	}
+	resp.Body.Close()
 	if resp.StatusCode != http.StatusOK {
 		return fmt.Errorf("unexpected status attempting to download update %d", resp.StatusCode)
 	}
-	resp.Body.Close()
-	etag := strings.Trim(resp.Header.Get("etag"), "\"")
-	if etag == "" {
-		slog.Debug("no etag detected, falling back to filename based dedup")
-		etag = "_"
-	}
 	filename := Installer
 	_, params, err := mime.ParseMediaType(resp.Header.Get("content-disposition"))
-	if err == nil {
+	if err == nil && params["filename"] != "" {
 		filename = params["filename"]
 	}

-	stageFilename := filepath.Join(UpdateStageDir, etag, filename)
+	stageFilename, err := updateStagePath(UpdateStageDir, resp.Header.Get("etag"), filename)
+	if err != nil {
+		return err
+	}

 	// Check to see if we already have it downloaded
 	_, err = os.Stat(stageFilename)
@@ -202,13 +202,14 @@ func (u *Updater) DownloadNewRelease(ctx context.Context, updateResp UpdateRespo
 		return fmt.Errorf("error checking update: %w", err)
 	}
 	defer resp.Body.Close()
-	etag = strings.Trim(resp.Header.Get("etag"), "\"")
-	if etag == "" {
-		slog.Debug("no etag detected, falling back to filename based dedup") // TODO probably can get rid of this redundant log
-		etag = "_"
+	if resp.StatusCode != http.StatusOK {
+		return fmt.Errorf("unexpected status attempting to download update %d", resp.StatusCode)
 	}

-	stageFilename = filepath.Join(UpdateStageDir, etag, filename)
+	stageFilename, err = updateStagePath(UpdateStageDir, resp.Header.Get("etag"), filename)
+	if err != nil {
+		return err
+	}

 	_, err = os.Stat(filepath.Dir(stageFilename))
 	if errors.Is(err, os.ErrNotExist) {
@@ -225,10 +226,13 @@ func (u *Updater) DownloadNewRelease(ctx context.Context, updateResp UpdateRespo
 	if err != nil {
 		return fmt.Errorf("write payload %s: %w", stageFilename, err)
 	}
-	defer fp.Close()
 	if n, err := fp.Write(payload); err != nil || n != len(payload) {
+		_ = fp.Close()
 		return fmt.Errorf("write payload %s: %d vs %d -- %w", stageFilename, n, len(payload), err)
 	}
+	if err := fp.Close(); err != nil {
+		return fmt.Errorf("close payload %s: %w", stageFilename, err)
+	}
 	slog.Info("new update downloaded " + stageFilename)

 	if err := VerifyDownload(); err != nil {
@@ -239,6 +243,61 @@ func (u *Updater) DownloadNewRelease(ctx context.Context, updateResp UpdateRespo
 	return nil
 }

+func updateStagePath(stageDir, etag, filename string) (string, error) {
+	filename, err := safeUpdateFilename(filename)
+	if err != nil {
+		return "", err
+	}
+
+	stageDir, err = filepath.Abs(stageDir)
+	if err != nil {
+		return "", fmt.Errorf("resolve update stage dir: %w", err)
+	}
+
+	stageFilename := filepath.Join(stageDir, updateStageETagDir(etag), filename)
+	if err := ensurePathInDir(stageDir, stageFilename); err != nil {
+		return "", err
+	}
+
+	return stageFilename, nil
+}
+
+func safeUpdateFilename(filename string) (string, error) {
+	filename = strings.TrimSpace(filename)
+	if filename == "" {
+		return "", errors.New("missing update filename")
+	}
+	if filename == "." || filename == ".." ||
+		filepath.IsAbs(filename) || path.IsAbs(filename) ||
+		strings.ContainsAny(filename, `/\:`) ||
+		filepath.Base(filename) != filename || path.Base(filename) != filename {
+		return "", fmt.Errorf("unsafe update filename %q", filename)
+	}
+	return filename, nil
+}
+
+func updateStageETagDir(etag string) string {
+	etag = strings.Trim(strings.TrimSpace(etag), "\"")
+	if etag == "" {
+		slog.Debug("no etag detected, falling back to filename based dedup")
+		return "_"
+	}
+
+	sum := sha256.Sum256([]byte(etag))
+	return hex.EncodeToString(sum[:])
+}
+
+func ensurePathInDir(dir, name string) error {
+	rel, err := filepath.Rel(dir, name)
+	if err != nil {
+		return fmt.Errorf("resolve update staging path: %w", err)
+	}
+	if rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator)) || filepath.IsAbs(rel) {
+		return fmt.Errorf("update staging path escapes stage dir: %s", name)
+	}
+	return nil
+}
+
 func cleanupOldDownloads(stageDir string) {
 	files, err := os.ReadDir(stageDir)
 	if err != nil && errors.Is(err, os.ErrNotExist) {
@@ -22,6 +22,15 @@ import (
 	"golang.org/x/sys/unix"
 )

+const updateArchiveRoot = "Ollama.app"
+
+type bundleEntryScope int
+
+const (
+	bundleEntryRelative bundleEntryScope = iota
+	bundleEntryWithArchiveRoot
+)
+
 var (
 	appBackupDir   string
 	SystemWidePath = "/Applications/Ollama.app"
@@ -167,8 +176,12 @@ func DoUpgrade(interactive bool) error {
 		}
 		name := s[1]
 		if strings.HasSuffix(name, "/") {
-			d := filepath.Join(BundlePath, name)
-			err := os.MkdirAll(d, 0o755)
+			d, err := bundleEntryPath(BundlePath, name, bundleEntryRelative)
+			if err != nil {
+				anyFailures = true
+				return err
+			}
+			err = os.MkdirAll(d, 0o755)
 			if err != nil {
 				anyFailures = true
 				return fmt.Errorf("failed to mkdir %s: %w", d, err)
@@ -181,30 +194,14 @@ func DoUpgrade(interactive bool) error {
 			continue
 		}

-		src, err := f.Open()
+		destName, err := bundleEntryPath(BundlePath, name, bundleEntryRelative)
 		if err != nil {
 			anyFailures = true
-			return fmt.Errorf("failed to open bundle file %s: %w", name, err)
+			return err
 		}
-		destName := filepath.Join(BundlePath, name)
-		// Verify directory first
-		d := filepath.Dir(destName)
-		if _, err := os.Stat(d); err != nil {
-			err := os.MkdirAll(d, 0o755)
-			if err != nil {
-				anyFailures = true
-				return fmt.Errorf("failed to mkdir %s: %w", d, err)
-			}
-		}
-		destFile, err := os.OpenFile(destName, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
-		if err != nil {
+		if err := extractBundleFile(f, destName, name); err != nil {
 			anyFailures = true
-			return fmt.Errorf("failed to open output file %s: %w", destName, err)
-		}
-		defer destFile.Close()
-		if _, err := io.Copy(destFile, src); err != nil {
-			anyFailures = true
-			return fmt.Errorf("failed to open extract file %s: %w", destName, err)
+			return err
 		}
 	}
 	for _, f := range links {
@@ -225,16 +222,24 @@ func DoUpgrade(interactive bool) error {
 			return err
 		}
 		link := string(buf)
-		if link[0] == '/' {
+		if link == "" {
+			anyFailures = true
+			return fmt.Errorf("bundle contains empty symlink %s", f.Name)
+		}
+		if filepath.IsAbs(link) {
 			anyFailures = true
 			return fmt.Errorf("bundle contains absolute symlink %s -> %s", f.Name, link)
 		}
-		// Don't allow links outside of Ollama.app
-		if strings.HasPrefix(filepath.Join(filepath.Dir(name), link), "..") {
+		if !validBundleLinkTarget(name, link, bundleEntryRelative) {
 			anyFailures = true
-			return fmt.Errorf("bundle contains link outside of contents %s -> %s", f.Name, link)
+			return fmt.Errorf("bundle contains invalid symlink %s -> %s", f.Name, link)
 		}
-		if err = os.Symlink(link, filepath.Join(BundlePath, name)); err != nil {
+		destName, err := bundleEntryPath(BundlePath, name, bundleEntryRelative)
+		if err != nil {
+			anyFailures = true
+			return err
+		}
+		if err = os.Symlink(link, destName); err != nil {
 			anyFailures = true
 			return err
 		}
@@ -282,8 +287,11 @@ func verifyDownload() error {
 	links := []*zip.File{}
 	for _, f := range r.File {
 		if strings.HasSuffix(f.Name, "/") {
-			d := filepath.Join(dir, f.Name)
-			err := os.MkdirAll(d, 0o755)
+			d, err := bundleEntryPath(dir, f.Name, bundleEntryWithArchiveRoot)
+			if err != nil {
+				return err
+			}
+			err = os.MkdirAll(d, 0o755)
 			if err != nil {
 				return fmt.Errorf("failed to mkdir %s: %w", d, err)
 			}
@@ -294,26 +302,12 @@ func verifyDownload() error {
 			links = append(links, f)
 			continue
 		}
-		src, err := f.Open()
+		destName, err := bundleEntryPath(dir, f.Name, bundleEntryWithArchiveRoot)
 		if err != nil {
-			return fmt.Errorf("failed to open bundle file %s: %w", f.Name, err)
+			return err
 		}
-		destName := filepath.Join(dir, f.Name)
-		// Verify directory first
-		d := filepath.Dir(destName)
-		if _, err := os.Stat(d); err != nil {
-			err := os.MkdirAll(d, 0o755)
-			if err != nil {
-				return fmt.Errorf("failed to mkdir %s: %w", d, err)
-			}
-		}
-		destFile, err := os.OpenFile(destName, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
-		if err != nil {
-			return fmt.Errorf("failed to open output file %s: %w", destName, err)
-		}
-		defer destFile.Close()
-		if _, err := io.Copy(destFile, src); err != nil {
-			return fmt.Errorf("failed to open extract file %s: %w", destName, err)
+		if err := extractBundleFile(f, destName, f.Name); err != nil {
+			return err
 		}
 	}
 	for _, f := range links {
@@ -326,13 +320,20 @@ func verifyDownload() error {
 			return err
 		}
 		link := string(buf)
-		if link[0] == '/' {
+		if link == "" {
+			return fmt.Errorf("bundle contains empty symlink %s", f.Name)
+		}
+		if filepath.IsAbs(link) {
 			return fmt.Errorf("bundle contains absolute symlink %s -> %s", f.Name, link)
 		}
-		if strings.HasPrefix(filepath.Join(filepath.Dir(f.Name), link), "..") {
-			return fmt.Errorf("bundle contains link outside of contents %s -> %s", f.Name, link)
+		if !validBundleLinkTarget(f.Name, link, bundleEntryWithArchiveRoot) {
+			return fmt.Errorf("bundle contains invalid symlink %s -> %s", f.Name, link)
 		}
-		if err = os.Symlink(link, filepath.Join(dir, f.Name)); err != nil {
+		destName, err := bundleEntryPath(dir, f.Name, bundleEntryWithArchiveRoot)
+		if err != nil {
+			return err
+		}
+		if err = os.Symlink(link, destName); err != nil {
 			return err
 		}
 	}
@@ -343,6 +344,53 @@ func verifyDownload() error {
 	return nil
 }

+func bundleEntryPath(root, name string, scope bundleEntryScope) (string, error) {
+	cleanName := filepath.Clean(filepath.FromSlash(name))
+	if !filepath.IsLocal(cleanName) {
+		return "", fmt.Errorf("bundle contains invalid path: %s", name)
+	}
+	if scope == bundleEntryWithArchiveRoot && cleanName != updateArchiveRoot &&
+		!strings.HasPrefix(cleanName, updateArchiveRoot+string(os.PathSeparator)) {
+		return "", fmt.Errorf("bundle contains invalid path: %s", name)
+	}
+	return filepath.Join(root, cleanName), nil
+}
+
+func extractBundleFile(f *zip.File, destName, name string) error {
+	src, err := f.Open()
+	if err != nil {
+		return fmt.Errorf("failed to open bundle file %s: %w", name, err)
+	}
+	defer src.Close()
+
+	d := filepath.Dir(destName)
+	if _, err := os.Stat(d); err != nil {
+		if err := os.MkdirAll(d, 0o755); err != nil {
+			return fmt.Errorf("failed to mkdir %s: %w", d, err)
+		}
+	}
+
+	destFile, err := os.OpenFile(destName, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+	if err != nil {
+		return fmt.Errorf("failed to open output file %s: %w", destName, err)
+	}
+	defer destFile.Close()
+
+	if _, err := io.Copy(destFile, src); err != nil {
+		return fmt.Errorf("failed to open extract file %s: %w", destName, err)
+	}
+	return nil
+}
+
+func validBundleLinkTarget(name, link string, scope bundleEntryScope) bool {
+	cleanTarget := filepath.Clean(filepath.Join(filepath.Dir(filepath.FromSlash(name)), filepath.FromSlash(link)))
+	if !filepath.IsLocal(cleanTarget) {
+		return false
+	}
+	return scope == bundleEntryRelative || cleanTarget == updateArchiveRoot ||
+		strings.HasPrefix(cleanTarget, updateArchiveRoot+string(os.PathSeparator))
+}
+
 // If we detect an upgrade bundle, attempt to upgrade at startup
 func DoUpgradeAtStartup() error {
 	bundle := getStagedUpdate()
@@ -2,6 +2,7 @@ package updater

 import (
 	"archive/zip"
+	"errors"
 	"io/fs"
 	"os"
 	"path/filepath"
@@ -146,6 +147,46 @@ func TestDoUpgrade(t *testing.T) {
 	}
 }

+func TestDoUpgradeRejectsInvalidBundlePath(t *testing.T) {
+	tmpDir := t.TempDir()
+	BundlePath = filepath.Join(tmpDir, "Ollama.app")
+	appBackupDir = filepath.Join(tmpDir, "backup")
+	UpdateStageDir = filepath.Join(tmpDir, "updates")
+	UpgradeMarkerFile = filepath.Join(tmpDir, "upgraded")
+	bundle := filepath.Join(UpdateStageDir, "foo", "ollama-darwin.zip")
+	invalidTarget := filepath.Join(tmpDir, "invalid-entry")
+
+	if err := os.MkdirAll(filepath.Join(BundlePath, "Contents", "MacOS"), 0o755); err != nil {
+		t.Fatal("failed to create empty dirs")
+	}
+	if err := os.WriteFile(filepath.Join(BundlePath, "Contents", "MacOS", "Ollama"), []byte("old app"), 0o755); err != nil {
+		t.Fatal("failed to create old app")
+	}
+	if err := os.MkdirAll(filepath.Dir(bundle), 0o755); err != nil {
+		t.Fatal("failed to create empty dirs")
+	}
+	if err := zipCreationHelper(bundle, []testPayload{{
+		Name: "Ollama.app/../invalid-entry",
+		Body: []byte("payload"),
+	}}); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := DoUpgrade(false); err == nil {
+		t.Fatal("expected failure with invalid bundle path")
+	} else if !strings.Contains(err.Error(), "bundle contains invalid path") {
+		t.Fatalf("unexpected error with invalid bundle path: %s", err)
+	}
+	if _, err := os.Stat(invalidTarget); err == nil {
+		t.Fatalf("invalid bundle path wrote %s", invalidTarget)
+	} else if !errors.Is(err, os.ErrNotExist) {
+		t.Fatalf("unexpected stat error for %s: %s", invalidTarget, err)
+	}
+	if _, err := os.Stat(filepath.Join(BundlePath, "Contents", "MacOS", "Ollama")); err != nil {
+		t.Fatalf("old app was not restored: %s", err)
+	}
+}
+
 func TestDoUpgradeAtStartup(t *testing.T) {
 	tmpDir := t.TempDir()
 	BundlePath = filepath.Join(tmpDir, "Ollama.app")
@@ -203,7 +244,7 @@ func TestVerifyDownloadFailures(t *testing.T) {
 		in       []testPayload
 		expected string
 	}{
-		{"breakout", []testPayload{
+		{"invalid symlink target", []testPayload{
 			{
 				Name: "Ollama.app/",
 				Body: []byte{},
@@ -212,15 +253,34 @@ func TestVerifyDownloadFailures(t *testing.T) {
 				Body: []byte("cli payload here"),
 			}, {
 				Name: "Ollama.app/Contents/MacOS/Ollama",
-				Body: []byte("../../../../breakout"),
+				Body: []byte("../../../../invalid-target"),
 				Mode: os.ModeSymlink,
 			},
-		}, "bundle contains link outside"},
+		}, "bundle contains invalid symlink"},
+		{"invalid archive symlink target", []testPayload{
+			{
+				Name: "Ollama.app/Contents/MacOS/Ollama",
+				Body: []byte("../../../invalid-target"),
+				Mode: os.ModeSymlink,
+			},
+		}, "bundle contains invalid symlink"},
 		{"absolute", []testPayload{{
 			Name: "Ollama.app/Contents/MacOS/Ollama",
 			Body: []byte("/etc/foo"),
 			Mode: os.ModeSymlink,
 		}}, "bundle contains absolute"},
+		{"invalid relative file", []testPayload{{
+			Name: "Ollama.app/../invalid-entry",
+			Body: []byte("payload"),
+		}}, "bundle contains invalid path"},
+		{"invalid relative directory", []testPayload{{
+			Name: "Ollama.app/../invalid-entry/",
+			Body: []byte{},
+		}}, "bundle contains invalid path"},
+		{"absolute file", []testPayload{{
+			Name: filepath.Join(tmpDir, "invalid-entry"),
+			Body: []byte("payload"),
+		}}, "bundle contains invalid path"},
 		{"missing", []testPayload{{
 			Name: "Ollama.app/Contents/MacOS/Ollama",
 			Body: []byte("../nothere"),
@@ -242,6 +302,11 @@ func TestVerifyDownloadFailures(t *testing.T) {
 			if err == nil || !strings.Contains(err.Error(), tt.expected) {
 				t.Fatalf("expected \"%s\" got %s", tt.expected, err)
 			}
+			if _, err := os.Stat(filepath.Join(tmpDir, "invalid-entry")); err == nil {
+				t.Fatal("invalid bundle path wrote unexpected file")
+			} else if !errors.Is(err, os.ErrNotExist) {
+				t.Fatalf("unexpected stat error for invalid file: %s", err)
+			}
 		})
 	}
 }
@@ -0,0 +1,127 @@
+//go:build (windows || darwin) && updater_live
+
+package updater
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/app/store"
+	"github.com/ollama/ollama/app/version"
+)
+
+// TestLiveAppUpdate exercises the production update endpoint and downloads the
+// current OS update artifact. It is intentionally excluded from normal test
+// runs because it depends on ollama.com and downloads a release artifact.
+//
+// Run with:
+//
+//	go test -tags updater_live -run TestLiveAppUpdate ./app/updater
+func TestLiveAppUpdate(t *testing.T) {
+	const spoofedVersion = "0.20.0"
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute)
+	defer cancel()
+
+	oldUpdateStageDir := UpdateStageDir
+	oldUpdateDownloaded := UpdateDownloaded
+	oldVerifyDownload := VerifyDownload
+	oldVersion := version.Version
+	defer func() {
+		UpdateStageDir = oldUpdateStageDir
+		UpdateDownloaded = oldUpdateDownloaded
+		VerifyDownload = oldVerifyDownload
+		version.Version = oldVersion
+	}()
+
+	version.Version = spoofedVersion
+
+	expectedFilename := ""
+	switch runtime.GOOS {
+	case "windows":
+		t.Setenv("LOCALAPPDATA", t.TempDir())
+		expectedFilename = "OllamaSetup.exe"
+	case "darwin":
+		expectedFilename = "Ollama-darwin.zip"
+	default:
+		t.Fatalf("unsupported updater live test OS %q", runtime.GOOS)
+	}
+
+	UpdateStageDir = filepath.Join(t.TempDir(), "updates")
+	UpdateDownloaded = false
+	verifyCalled := false
+	VerifyDownload = func() error {
+		verifyCalled = true
+		return verifyDownload()
+	}
+
+	updater := &Updater{Store: &store.Store{DBPath: filepath.Join(t.TempDir(), "db.sqlite")}}
+	defer updater.Store.Close()
+
+	available, updateResp := updater.checkForUpdate(ctx)
+	if !available {
+		t.Fatalf("expected production update check to offer an update for spoofed version %s", spoofedVersion)
+	}
+	if updateResp.UpdateURL == "" {
+		t.Fatal("production update response did not include a download URL")
+	}
+	t.Logf("production update version=%q url=%q", updateResp.UpdateVersion, updateResp.UpdateURL)
+
+	if err := updater.DownloadNewRelease(ctx, updateResp); err != nil {
+		t.Fatalf("download production update: %v", err)
+	}
+
+	staged := getStagedUpdate()
+	if staged == "" {
+		t.Fatal("production update was not staged")
+	}
+	t.Logf("staged production update at %s", staged)
+
+	assertPathInsideDir(t, UpdateStageDir, staged)
+	if filepath.Base(staged) != expectedFilename {
+		t.Fatalf("expected staged %s update filename to be %q, got %q", runtime.GOOS, expectedFilename, filepath.Base(staged))
+	}
+	expectedExt := filepath.Ext(expectedFilename)
+	if filepath.Ext(staged) != expectedExt {
+		t.Fatalf("expected staged %s update to be a %s artifact, got %s", runtime.GOOS, expectedExt, staged)
+	}
+
+	info, err := os.Stat(staged)
+	if err != nil {
+		t.Fatalf("stat staged update: %v", err)
+	}
+	if info.Size() == 0 {
+		t.Fatal("staged production update is empty")
+	}
+
+	if !verifyCalled {
+		t.Fatal("DownloadNewRelease did not call VerifyDownload")
+	}
+	t.Logf("production updater download path verified staged %s update", runtime.GOOS)
+}
+
+func assertPathInsideDir(t *testing.T, dir, name string) {
+	t.Helper()
+
+	dir, err := filepath.Abs(dir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	name, err = filepath.Abs(name)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	rel, err := filepath.Rel(dir, name)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator)) || filepath.IsAbs(rel) {
+		t.Fatalf("staged update escaped update stage dir: %s", name)
+	}
+}
--- a/Show More
+++ b/Show More
				`@@ -0,0 +1 @@`
				`<svg height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Claude Code</title><path clip-rule="evenodd" d="M20.998 10.949H24v3.102h-3v3.028h-1.487V20H18v-2.921h-1.487V20H15v-2.921H9V20H7.488v-2.921H6V20H4.487v-2.921H3V14.05H0V10.95h3V5h17.998v5.949zM6 10.949h1.488V8.102H6v2.847zm10.51 0H18V8.102h-1.49v2.847z" fill="#D97757" fill-rule="evenodd"></path></svg>`
				`@@ -0,0 +1 @@`
				<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 320 320"><path fill="#fff" d="m297.06 130.97c7.26-21.79 4.76-45.66-6.85-65.48-17.46-30.4-52.56-46.04-86.84-38.68-15.25-17.18-37.16-26.95-60.13-26.81-35.04-.08-66.13 22.48-76.91 55.82-22.51 4.61-41.94 18.7-53.31 38.67-17.59 30.32-13.58 68.54 9.92 94.54-7.26 21.79-4.76 45.66 6.85 65.48 17.46 30.4 52.56 46.04 86.84 38.68 15.24 17.18 37.16 26.95 60.13 26.8 35.06.09 66.16-22.49 76.94-55.86 22.51-4.61 41.94-18.7 53.31-38.67 17.57-30.32 13.55-68.51-9.94-94.51zm-120.28 168.11c-14.03.02-27.62-4.89-38.39-13.88.49-.26 1.34-.73 1.89-1.07l63.72-36.8c3.26-1.85 5.26-5.32 5.24-9.07v-89.83l26.93 15.55c.29.14.48.42.52.74v74.39c-.04 33.08-26.83 59.9-59.91 59.97zm-128.84-55.03c-7.03-12.14-9.56-26.37-7.15-40.18.47.28 1.3.79 1.89 1.13l63.72 36.8c3.23 1.89 7.23 1.89 10.47 0l77.79-44.92v31.1c.02.32-.13.63-.38.83l-64.41 37.19c-28.69 16.52-65.33 6.7-81.92-21.95zm-16.77-139.09c7-12.16 18.05-21.46 31.21-26.29 0 .55-.03 1.52-.03 2.2v73.61c-.02 3.74 1.98 7.21 5.23 9.06l77.79 44.91-26.93 15.55c-.27.18-.61.21-.91.08l-64.42-37.22c-28.63-16.58-38.45-53.21-21.95-81.89zm221.26 51.49-77.79-44.92 26.93-15.54c.27-.18.61-.21.91-.08l64.42 37.19c28.68 16.57 38.51 53.26 21.94 81.94-7.01 12.14-18.05 21.44-31.2 26.28v-75.81c.03-3.74-1.96-7.2-5.2-9.06zm26.8-40.34c-.47-.29-1.3-.79-1.89-1.13l-63.72-36.8c-3.23-1.89-7.23-1.89-10.47 0l-77.79 44.92v-31.1c-.02-.32.13-.63.38-.83l64.41-37.16c28.69-16.55 65.37-6.7 81.91 22 6.99 12.12 9.52 26.31 7.15 40.1zm-168.51 55.43-26.94-15.55c-.29-.14-.48-.42-.52-.74v-74.39c.02-33.12 26.89-59.96 60.01-59.94 14.01 0 27.57 4.92 38.34 13.88-.49.26-1.33.73-1.89 1.07l-63.72 36.8c-3.26 1.85-5.26 5.31-5.24 9.06l-.04 89.79zm14.63-31.54 34.65-20.01 34.65 20v40.01l-34.65 20-34.65-20z"/></svg>