cmd/config: fix droid model handling and naming

- Remove unused strings import
- Fix Ollama model identification logic to properly detect Ollama models
- Simplify model ID naming scheme by removing [Ollama] suffix
- Improve type assertions for better type safety
- Remove redundant helper function isOllamaModelEntry

.github/ISSUE_TEMPLATE/10_bug_report.yml

@@ -13,7 +13,7 @@ body:
     id: logs
     attributes:
       label: Relevant log output
-      description: Please copy and paste any relevant log output. See [Troubleshooting Guide](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) for details.
+      description: Please copy and paste any relevant log output. See [Troubleshooting Guide](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.mdx#how-to-troubleshoot-issues) for details.
       render: shell
     validations:
       required: false

.github/workflows/release.yaml

@@ -68,6 +68,7 @@ jobs:
           name: bundles-darwin
           path: |
             dist/*.tgz
+            dist/*.tar.zst
             dist/*.zip
             dist/*.dmg
 
@@ -371,13 +372,17 @@ jobs:
           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 ;;
+              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/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 }}-${{ matrix.arch }}-rocm.tar.in ;;
@@ -392,13 +397,13 @@ jobs:
           done
       - run: |
           for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
-            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | pigz -9vc >$(basename ${ARCHIVE//.*/}.tgz);
+            tar c -C dist/${{ 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: |
-            *.tgz
+            *.tar.zst
 
   # Build each Docker variant (OS, arch, and flavor) separately. Using QEMU is unreliable and slower.
   docker-build-push:
@@ -531,7 +536,7 @@ jobs:
       - name: Upload release artifacts
         run: |
           pids=()
-          for payload in dist/*.txt dist/*.zip dist/*.tgz dist/*.exe dist/*.dmg ; do
+          for payload in dist/*.txt dist/*.zip dist/*.tgz dist/*.tar.zst dist/*.exe dist/*.dmg ; do
             echo "Uploading $payload"
             gh release upload ${GITHUB_REF_NAME} $payload --clobber &
             pids[$!]=$!

CMakeLists.txt

@@ -2,6 +2,22 @@ cmake_minimum_required(VERSION 3.21)
 
 project(Ollama C CXX)
 
+# Handle cross-compilation on macOS: when CMAKE_OSX_ARCHITECTURES is set to a
+# single architecture different from the host, override CMAKE_SYSTEM_PROCESSOR
+# to match. This is necessary because CMAKE_SYSTEM_PROCESSOR defaults to the
+# host architecture, but downstream projects (like MLX) use it to detect the
+# target architecture.
+if(CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES ";")
+    # Single architecture specified
+    if(CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" AND NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+        message(STATUS "Cross-compiling for x86_64: overriding CMAKE_SYSTEM_PROCESSOR from ${CMAKE_SYSTEM_PROCESSOR} to x86_64")
+        set(CMAKE_SYSTEM_PROCESSOR "x86_64")
+    elseif(CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" AND NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
+        message(STATUS "Cross-compiling for arm64: overriding CMAKE_SYSTEM_PROCESSOR from ${CMAKE_SYSTEM_PROCESSOR} to arm64")
+        set(CMAKE_SYSTEM_PROCESSOR "arm64")
+    endif()
+endif()
+
 include(CheckLanguage)
 include(GNUInstallDirs)
 
@@ -12,7 +28,7 @@ set(BUILD_SHARED_LIBS ON)
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_EXTENSIONS OFF)
+set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX Requires gnu++17 extensions to compile properly
 
 set(GGML_BUILD ON)
 set(GGML_SHARED ON)
@@ -32,9 +48,10 @@ if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
     set(GGML_CPU_ALL_VARIANTS ON)
 endif()
 
-if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
+if(APPLE)
     set(CMAKE_BUILD_RPATH "@loader_path")
     set(CMAKE_INSTALL_RPATH "@loader_path")
+    set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
 endif()
 
 set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
@@ -147,14 +164,56 @@ if(CMAKE_HIP_COMPILER)
     endif()
 endif()
 
-find_package(Vulkan)
-if(Vulkan_FOUND)
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
-    install(TARGETS ggml-vulkan
-        RUNTIME_DEPENDENCIES
-            PRE_INCLUDE_REGEXES vulkan
-            PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-    )
+if(NOT APPLE)
+    find_package(Vulkan)
+    if(Vulkan_FOUND)
+        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
+        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/ml/backend/mlx)
+
+    # Find CUDA toolkit if MLX is built with CUDA support
+    find_package(CUDAToolkit)
+
+    install(TARGETS mlx mlxc
+        RUNTIME_DEPENDENCIES
+            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
+            PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc nvrtc-builtins cudnn nccl openblas gfortran
+            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 ${OLLAMA_INSTALL_DIR}
+            COMPONENT MLX)
+    endif()
+
+    # Manually install cudart and cublas since they might not be picked up as direct dependencies
+    if(CUDAToolkit_FOUND)
+        file(GLOB CUDART_LIBS
+            "${CUDAToolkit_LIBRARY_DIR}/libcudart.so*"
+            "${CUDAToolkit_LIBRARY_DIR}/libcublas.so*")
+        if(CUDART_LIBS)
+            install(FILES ${CUDART_LIBS}
+                DESTINATION ${OLLAMA_INSTALL_DIR}
+                COMPONENT MLX)
+        endif()
+    endif()
+endif()

CMakePresets.json

@@ -41,7 +41,7 @@
       "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 2",
+        "CMAKE_CUDA_FLAGS": "-t 4",
         "OLLAMA_RUNNER_DIR": "cuda_v13"
       }
     },
@@ -83,6 +83,28 @@
       "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": {
+        "OLLAMA_RUNNER_DIR": "mlx_cuda_v13"
+      }
     }
   ],
   "buildPresets": [
@@ -140,6 +162,21 @@
       "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"
     }
   ]
 }

Dockerfile

@@ -32,7 +32,7 @@ ENV PATH=/${VULKANVERSION}/x86_64/bin:$PATH
 FROM --platform=linux/arm64 almalinux:8 AS base-arm64
 # install epel-release for ccache
 RUN yum install -y yum-utils epel-release \
-    && dnf install -y clang ccache \
+    && dnf install -y clang ccache git \
     && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
 ENV CC=clang CXX=clang++
 
@@ -131,8 +131,32 @@ COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'Vulkan' \
         && cmake --build --parallel --preset 'Vulkan' \
-        && cmake --install build --component Vulkan --strip --parallel 8 
+        && cmake --install build --component Vulkan --strip --parallel 8
 
+FROM base AS mlx
+ARG CUDA13VERSION=13.0
+RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-} \
+    && dnf install -y openblas-devel lapack-devel \
+    && dnf install -y libcudnn9-cuda-13 libcudnn9-devel-cuda-13 \
+    && dnf install -y libnccl libnccl-devel
+ENV PATH=/usr/local/cuda-13/bin:$PATH
+ENV BLAS_INCLUDE_DIRS=/usr/include/openblas
+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"
+ARG PARALLEL
+WORKDIR /go/src/github.com/ollama/ollama
+COPY CMakeLists.txt CMakePresets.json .
+COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+COPY x/ml/backend/mlx x/ml/backend/mlx
+COPY go.mod go.sum .
+COPY MLX_VERSION .
+RUN curl -fsSL https://golang.org/dl/go$(awk '/^go/ { print $2 }' go.mod).linux-$(case $(uname -m) in x86_64) echo amd64 ;; aarch64) echo arm64 ;; esac).tar.gz | tar xz -C /usr/local
+ENV PATH=/usr/local/go/bin:$PATH
+RUN go mod download
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'MLX CUDA 13' -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas \
+        && cmake --build --parallel ${PARALLEL} --preset 'MLX CUDA 13' \
+        && cmake --install build --component MLX --strip --parallel ${PARALLEL}
 
 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -141,18 +165,21 @@ RUN curl -fsSL https://golang.org/dl/go$(awk '/^go/ { print $2 }' go.mod).linux-
 ENV PATH=/usr/local/go/bin:$PATH
 RUN go mod download
 COPY . .
+# Clone mlx-c headers for CGO (version from MLX_VERSION file)
+RUN git clone --depth 1 --branch "$(cat MLX_VERSION)" https://github.com/ml-explore/mlx-c.git build/_deps/mlx-c-src
 ARG GOFLAGS="'-ldflags=-w -s'"
 ENV CGO_ENABLED=1
-ARG CGO_CFLAGS
+ENV CGO_CFLAGS="-I/go/src/github.com/ollama/ollama/build/_deps/mlx-c-src"
 ARG CGO_CXXFLAGS
 RUN --mount=type=cache,target=/root/.cache/go-build \
-    go build -trimpath -buildmode=pie -o /bin/ollama .
+    go build -tags mlx -trimpath -buildmode=pie -o /bin/ollama .
 
 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=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/
@@ -171,7 +198,7 @@ COPY --from=build /bin/ollama /bin/ollama
 
 FROM ubuntu:24.04
 RUN apt-get update \
-    && apt-get install -y ca-certificates libvulkan1 \
+    && apt-get install -y ca-certificates libvulkan1 libopenblas0 \
     && apt-get clean \
     && rm -rf /var/lib/apt/lists/*
 COPY --from=archive /bin /usr/bin

MLX_VERSION

@@ -0,0 +1 @@
+v0.4.1

README.md

@@ -48,7 +48,7 @@ ollama run gemma3
 
 ## Model library
 
-Ollama supports a list of models available on [ollama.com/library](https://ollama.com/library 'ollama model library')
+Ollama supports a list of models available on [ollama.com/library](https://ollama.com/library "ollama model library")
 
 Here are some example models that can be downloaded:
 
@@ -79,7 +79,7 @@ Here are some example models that can be downloaded:
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`           |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored`   |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`               |
-| Granite-3.3         | 8B         | 4.9GB | `ollama run granite3.3`          |
+| Granite-3.3        | 8B         | 4.9GB | `ollama run granite3.3`          |
 
 > [!NOTE]
 > You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
@@ -260,6 +260,38 @@ Finally, in a separate shell, run a model:
 ./ollama run llama3.2
 ```
 
+## Building with MLX (experimental)
+
+First build the MLX libraries:
+
+```shell
+cmake --preset MLX
+cmake --build --preset MLX --parallel
+cmake --install build --component MLX
+```
+
+When building with the `-tags mlx` flag, the main `ollama` binary includes MLX support for experimental features like image generation:
+
+```shell
+go build -tags mlx .
+```
+
+Finally, start the server:
+
+```
+./ollama serve
+```
+
+### Building MLX with CUDA
+
+When building with CUDA, use the preset "MLX CUDA 13" or "MLX CUDA 12" to enable CUDA with default architectures:
+
+```shell
+cmake --preset 'MLX CUDA 13'
+cmake --build --preset 'MLX CUDA 13' --parallel
+cmake --install build --component MLX
+```
+
 ## REST API
 
 Ollama has a REST API for running and managing models.
@@ -290,6 +322,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### Web & Desktop
 
+- [Onyx](https://github.com/onyx-dot-app/onyx)
 - [Open WebUI](https://github.com/open-webui/open-webui)
 - [SwiftChat (macOS with ReactNative)](https://github.com/aws-samples/swift-chat)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
@@ -421,7 +454,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [AppFlowy](https://github.com/AppFlowy-IO/AppFlowy) (AI collaborative workspace with Ollama, cross-platform and self-hostable)
 - [Lumina](https://github.com/cushydigit/lumina.git) (A lightweight, minimal React.js frontend for interacting with Ollama servers)
 - [Tiny Notepad](https://pypi.org/project/tiny-notepad) (A lightweight, notepad-like interface to chat with ollama available on PyPI)
-- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.) 
+- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.)
 - [GPTranslate](https://github.com/philberndt/GPTranslate) (A fast and lightweight, AI powered desktop translation application written with Rust and Tauri. Features real-time translation with OpenAI/Azure/Ollama.)
 - [ollama launcher](https://github.com/NGC13009/ollama-launcher) (A launcher for Ollama, aiming to provide users with convenient functions such as ollama server launching, management, or configuration.)
 - [ai-hub](https://github.com/Aj-Seven/ai-hub) (AI Hub supports multiple models via API keys and Chat support via Ollama API.)
@@ -493,7 +526,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Database
 
 - [pgai](https://github.com/timescale/pgai) - PostgreSQL as a vector database (Create and search embeddings from Ollama models using pgvector)
-   - [Get started guide](https://github.com/timescale/pgai/blob/main/docs/vectorizer-quick-start.md)
+  - [Get started guide](https://github.com/timescale/pgai/blob/main/docs/vectorizer-quick-start.md)
 - [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md) (Connects Ollama models with nearly 200 data platforms and apps)
 - [chromem-go](https://github.com/philippgille/chromem-go/blob/v0.5.0/embed_ollama.go) with [example](https://github.com/philippgille/chromem-go/tree/v0.5.0/examples/rag-wikipedia-ollama)
 - [Kangaroo](https://github.com/dbkangaroo/kangaroo) (AI-powered SQL client and admin tool for popular databases)
@@ -636,6 +669,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [llama.cpp](https://github.com/ggml-org/llama.cpp) project founded by Georgi Gerganov.
 
 ### Observability
+
 - [Opik](https://www.comet.com/docs/opik/cookbook/ollama) is an open-source platform to debug, evaluate, and monitor your LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards. Opik supports native integration to Ollama.
 - [Lunary](https://lunary.ai/docs/integrations/ollama) is the leading open-source LLM observability platform. It provides a variety of enterprise-grade features such as real-time analytics, prompt templates management, PII masking, and comprehensive agent tracing.
 - [OpenLIT](https://github.com/openlit/openlit) is an OpenTelemetry-native tool for monitoring Ollama Applications & GPUs using traces and metrics.
@@ -644,4 +678,5 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [MLflow Tracing](https://mlflow.org/docs/latest/llms/tracing/index.html#automatic-tracing) is an open source LLM observability tool with a convenient API to log and visualize traces, making it easy to debug and evaluate GenAI applications.
 
 ### Security
+
 - [Ollama Fortress](https://github.com/ParisNeo/ollama_proxy_server)

anthropic/anthropic.go

@@ -0,0 +1,778 @@
+package anthropic
+
+import (
+	"crypto/rand"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Error types matching Anthropic API
+type Error struct {
+	Type    string `json:"type"`
+	Message string `json:"message"`
+}
+
+type ErrorResponse struct {
+	Type      string `json:"type"` // always "error"
+	Error     Error  `json:"error"`
+	RequestID string `json:"request_id,omitempty"`
+}
+
+// NewError creates a new ErrorResponse with the appropriate error type based on HTTP status code
+func NewError(code int, message string) ErrorResponse {
+	var etype string
+	switch code {
+	case http.StatusBadRequest:
+		etype = "invalid_request_error"
+	case http.StatusUnauthorized:
+		etype = "authentication_error"
+	case http.StatusForbidden:
+		etype = "permission_error"
+	case http.StatusNotFound:
+		etype = "not_found_error"
+	case http.StatusTooManyRequests:
+		etype = "rate_limit_error"
+	case http.StatusServiceUnavailable, 529:
+		etype = "overloaded_error"
+	default:
+		etype = "api_error"
+	}
+
+	return ErrorResponse{
+		Type:      "error",
+		Error:     Error{Type: etype, Message: message},
+		RequestID: generateID("req"),
+	}
+}
+
+// Request types
+
+// MessagesRequest represents an Anthropic Messages API request
+type MessagesRequest struct {
+	Model         string          `json:"model"`
+	MaxTokens     int             `json:"max_tokens"`
+	Messages      []MessageParam  `json:"messages"`
+	System        any             `json:"system,omitempty"` // string or []ContentBlock
+	Stream        bool            `json:"stream,omitempty"`
+	Temperature   *float64        `json:"temperature,omitempty"`
+	TopP          *float64        `json:"top_p,omitempty"`
+	TopK          *int            `json:"top_k,omitempty"`
+	StopSequences []string        `json:"stop_sequences,omitempty"`
+	Tools         []Tool          `json:"tools,omitempty"`
+	ToolChoice    *ToolChoice     `json:"tool_choice,omitempty"`
+	Thinking      *ThinkingConfig `json:"thinking,omitempty"`
+	Metadata      *Metadata       `json:"metadata,omitempty"`
+}
+
+// MessageParam represents a message in the request
+type MessageParam struct {
+	Role    string `json:"role"`    // "user" or "assistant"
+	Content any    `json:"content"` // string or []ContentBlock
+}
+
+// ContentBlock represents a content block in a message.
+// Text and Thinking use pointers so they serialize as the field being present (even if empty)
+// only when set, which is required for SDK streaming accumulation.
+type ContentBlock struct {
+	Type string `json:"type"` // text, image, tool_use, tool_result, thinking
+
+	// For text blocks - pointer so field only appears when set (SDK requires it for accumulation)
+	Text *string `json:"text,omitempty"`
+
+	// For image blocks
+	Source *ImageSource `json:"source,omitempty"`
+
+	// For tool_use blocks
+	ID    string `json:"id,omitempty"`
+	Name  string `json:"name,omitempty"`
+	Input any    `json:"input,omitempty"`
+
+	// For tool_result blocks
+	ToolUseID string `json:"tool_use_id,omitempty"`
+	Content   any    `json:"content,omitempty"` // string or []ContentBlock
+	IsError   bool   `json:"is_error,omitempty"`
+
+	// For thinking blocks - pointer so field only appears when set (SDK requires it for accumulation)
+	Thinking  *string `json:"thinking,omitempty"`
+	Signature string  `json:"signature,omitempty"`
+}
+
+// ImageSource represents the source of an image
+type ImageSource struct {
+	Type      string `json:"type"` // "base64" or "url"
+	MediaType string `json:"media_type,omitempty"`
+	Data      string `json:"data,omitempty"`
+	URL       string `json:"url,omitempty"`
+}
+
+// Tool represents a tool definition
+type Tool struct {
+	Type        string          `json:"type,omitempty"` // "custom" for user-defined tools
+	Name        string          `json:"name"`
+	Description string          `json:"description,omitempty"`
+	InputSchema json.RawMessage `json:"input_schema,omitempty"`
+}
+
+// ToolChoice controls how the model uses tools
+type ToolChoice struct {
+	Type                   string `json:"type"` // "auto", "any", "tool", "none"
+	Name                   string `json:"name,omitempty"`
+	DisableParallelToolUse bool   `json:"disable_parallel_tool_use,omitempty"`
+}
+
+// ThinkingConfig controls extended thinking
+type ThinkingConfig struct {
+	Type         string `json:"type"` // "enabled" or "disabled"
+	BudgetTokens int    `json:"budget_tokens,omitempty"`
+}
+
+// Metadata for the request
+type Metadata struct {
+	UserID string `json:"user_id,omitempty"`
+}
+
+// Response types
+
+// MessagesResponse represents an Anthropic Messages API response
+type MessagesResponse struct {
+	ID           string         `json:"id"`
+	Type         string         `json:"type"` // "message"
+	Role         string         `json:"role"` // "assistant"
+	Model        string         `json:"model"`
+	Content      []ContentBlock `json:"content"`
+	StopReason   string         `json:"stop_reason,omitempty"`
+	StopSequence string         `json:"stop_sequence,omitempty"`
+	Usage        Usage          `json:"usage"`
+}
+
+// Usage contains token usage information
+type Usage struct {
+	InputTokens  int `json:"input_tokens"`
+	OutputTokens int `json:"output_tokens"`
+}
+
+// Streaming event types
+
+// MessageStartEvent is sent at the start of streaming
+type MessageStartEvent struct {
+	Type    string           `json:"type"` // "message_start"
+	Message MessagesResponse `json:"message"`
+}
+
+// ContentBlockStartEvent signals the start of a content block
+type ContentBlockStartEvent struct {
+	Type         string       `json:"type"` // "content_block_start"
+	Index        int          `json:"index"`
+	ContentBlock ContentBlock `json:"content_block"`
+}
+
+// ContentBlockDeltaEvent contains incremental content updates
+type ContentBlockDeltaEvent struct {
+	Type  string `json:"type"` // "content_block_delta"
+	Index int    `json:"index"`
+	Delta Delta  `json:"delta"`
+}
+
+// Delta represents an incremental update
+type Delta struct {
+	Type        string `json:"type"` // "text_delta", "input_json_delta", "thinking_delta", "signature_delta"
+	Text        string `json:"text,omitempty"`
+	PartialJSON string `json:"partial_json,omitempty"`
+	Thinking    string `json:"thinking,omitempty"`
+	Signature   string `json:"signature,omitempty"`
+}
+
+// ContentBlockStopEvent signals the end of a content block
+type ContentBlockStopEvent struct {
+	Type  string `json:"type"` // "content_block_stop"
+	Index int    `json:"index"`
+}
+
+// MessageDeltaEvent contains updates to the message
+type MessageDeltaEvent struct {
+	Type  string       `json:"type"` // "message_delta"
+	Delta MessageDelta `json:"delta"`
+	Usage DeltaUsage   `json:"usage"`
+}
+
+// MessageDelta contains stop information
+type MessageDelta struct {
+	StopReason   string `json:"stop_reason,omitempty"`
+	StopSequence string `json:"stop_sequence,omitempty"`
+}
+
+// DeltaUsage contains cumulative token usage
+type DeltaUsage struct {
+	OutputTokens int `json:"output_tokens"`
+}
+
+// MessageStopEvent signals the end of the message
+type MessageStopEvent struct {
+	Type string `json:"type"` // "message_stop"
+}
+
+// PingEvent is a keepalive event
+type PingEvent struct {
+	Type string `json:"type"` // "ping"
+}
+
+// StreamErrorEvent is an error during streaming
+type StreamErrorEvent struct {
+	Type  string `json:"type"` // "error"
+	Error Error  `json:"error"`
+}
+
+// FromMessagesRequest converts an Anthropic MessagesRequest to an Ollama api.ChatRequest
+func FromMessagesRequest(r MessagesRequest) (*api.ChatRequest, error) {
+	var messages []api.Message
+
+	if r.System != nil {
+		switch sys := r.System.(type) {
+		case string:
+			if sys != "" {
+				messages = append(messages, api.Message{Role: "system", Content: sys})
+			}
+		case []any:
+			// System can be an array of content blocks
+			var content strings.Builder
+			for _, block := range sys {
+				if blockMap, ok := block.(map[string]any); ok {
+					if blockMap["type"] == "text" {
+						if text, ok := blockMap["text"].(string); ok {
+							content.WriteString(text)
+						}
+					}
+				}
+			}
+			if content.Len() > 0 {
+				messages = append(messages, api.Message{Role: "system", Content: content.String()})
+			}
+		}
+	}
+
+	for _, msg := range r.Messages {
+		converted, err := convertMessage(msg)
+		if err != nil {
+			return nil, err
+		}
+		messages = append(messages, converted...)
+	}
+
+	options := make(map[string]any)
+
+	options["num_predict"] = r.MaxTokens
+
+	if r.Temperature != nil {
+		options["temperature"] = *r.Temperature
+	}
+
+	if r.TopP != nil {
+		options["top_p"] = *r.TopP
+	}
+
+	if r.TopK != nil {
+		options["top_k"] = *r.TopK
+	}
+
+	if len(r.StopSequences) > 0 {
+		options["stop"] = r.StopSequences
+	}
+
+	var tools api.Tools
+	for _, t := range r.Tools {
+		tool, err := convertTool(t)
+		if err != nil {
+			return nil, err
+		}
+		tools = append(tools, tool)
+	}
+
+	var think *api.ThinkValue
+	if r.Thinking != nil && r.Thinking.Type == "enabled" {
+		think = &api.ThinkValue{Value: true}
+	}
+
+	stream := r.Stream
+
+	return &api.ChatRequest{
+		Model:    r.Model,
+		Messages: messages,
+		Options:  options,
+		Stream:   &stream,
+		Tools:    tools,
+		Think:    think,
+	}, nil
+}
+
+// convertMessage converts an Anthropic MessageParam to Ollama api.Message(s)
+func convertMessage(msg MessageParam) ([]api.Message, error) {
+	var messages []api.Message
+	role := strings.ToLower(msg.Role)
+
+	switch content := msg.Content.(type) {
+	case string:
+		messages = append(messages, api.Message{Role: role, Content: content})
+
+	case []any:
+		var textContent strings.Builder
+		var images []api.ImageData
+		var toolCalls []api.ToolCall
+		var thinking string
+		var toolResults []api.Message
+
+		for _, block := range content {
+			blockMap, ok := block.(map[string]any)
+			if !ok {
+				return nil, errors.New("invalid content block format")
+			}
+
+			blockType, _ := blockMap["type"].(string)
+
+			switch blockType {
+			case "text":
+				if text, ok := blockMap["text"].(string); ok {
+					textContent.WriteString(text)
+				}
+
+			case "image":
+				source, ok := blockMap["source"].(map[string]any)
+				if !ok {
+					return nil, errors.New("invalid image source")
+				}
+
+				sourceType, _ := source["type"].(string)
+				if sourceType == "base64" {
+					data, _ := source["data"].(string)
+					decoded, err := base64.StdEncoding.DecodeString(data)
+					if err != nil {
+						return nil, fmt.Errorf("invalid base64 image data: %w", err)
+					}
+					images = append(images, decoded)
+				} else {
+					return nil, fmt.Errorf("invalid image source type: %s. Only base64 images are supported.", sourceType)
+				}
+				// URL images would need to be fetched - skip for now
+
+			case "tool_use":
+				id, ok := blockMap["id"].(string)
+				if !ok {
+					return nil, errors.New("tool_use block missing required 'id' field")
+				}
+				name, ok := blockMap["name"].(string)
+				if !ok {
+					return nil, errors.New("tool_use block missing required 'name' field")
+				}
+				tc := api.ToolCall{
+					ID: id,
+					Function: api.ToolCallFunction{
+						Name: name,
+					},
+				}
+				if input, ok := blockMap["input"].(map[string]any); ok {
+					tc.Function.Arguments = mapToArgs(input)
+				}
+				toolCalls = append(toolCalls, tc)
+
+			case "tool_result":
+				toolUseID, _ := blockMap["tool_use_id"].(string)
+				var resultContent string
+
+				switch c := blockMap["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
+								}
+							}
+						}
+					}
+				}
+
+				toolResults = append(toolResults, api.Message{
+					Role:       "tool",
+					Content:    resultContent,
+					ToolCallID: toolUseID,
+				})
+
+			case "thinking":
+				if t, ok := blockMap["thinking"].(string); ok {
+					thinking = t
+				}
+			}
+		}
+
+		if textContent.Len() > 0 || len(images) > 0 || len(toolCalls) > 0 || thinking != "" {
+			m := api.Message{
+				Role:      role,
+				Content:   textContent.String(),
+				Images:    images,
+				ToolCalls: toolCalls,
+				Thinking:  thinking,
+			}
+			messages = append(messages, m)
+		}
+
+		// Add tool results as separate messages
+		messages = append(messages, toolResults...)
+
+	default:
+		return nil, fmt.Errorf("invalid message content type: %T", content)
+	}
+
+	return messages, nil
+}
+
+// convertTool converts an Anthropic Tool to an Ollama api.Tool
+func convertTool(t Tool) (api.Tool, error) {
+	var params api.ToolFunctionParameters
+	if len(t.InputSchema) > 0 {
+		if err := json.Unmarshal(t.InputSchema, &params); err != nil {
+			return api.Tool{}, fmt.Errorf("invalid input_schema for tool %q: %w", t.Name, err)
+		}
+	}
+
+	return api.Tool{
+		Type: "function",
+		Function: api.ToolFunction{
+			Name:        t.Name,
+			Description: t.Description,
+			Parameters:  params,
+		},
+	}, nil
+}
+
+// ToMessagesResponse converts an Ollama api.ChatResponse to an Anthropic MessagesResponse
+func ToMessagesResponse(id string, r api.ChatResponse) MessagesResponse {
+	var content []ContentBlock
+
+	if r.Message.Thinking != "" {
+		content = append(content, ContentBlock{
+			Type:     "thinking",
+			Thinking: ptr(r.Message.Thinking),
+		})
+	}
+
+	if r.Message.Content != "" {
+		content = append(content, ContentBlock{
+			Type: "text",
+			Text: ptr(r.Message.Content),
+		})
+	}
+
+	for _, tc := range r.Message.ToolCalls {
+		content = append(content, ContentBlock{
+			Type:  "tool_use",
+			ID:    tc.ID,
+			Name:  tc.Function.Name,
+			Input: tc.Function.Arguments,
+		})
+	}
+
+	stopReason := mapStopReason(r.DoneReason, len(r.Message.ToolCalls) > 0)
+
+	return MessagesResponse{
+		ID:         id,
+		Type:       "message",
+		Role:       "assistant",
+		Model:      r.Model,
+		Content:    content,
+		StopReason: stopReason,
+		Usage: Usage{
+			InputTokens:  r.Metrics.PromptEvalCount,
+			OutputTokens: r.Metrics.EvalCount,
+		},
+	}
+}
+
+// mapStopReason converts Ollama done_reason to Anthropic stop_reason
+func mapStopReason(reason string, hasToolCalls bool) string {
+	if hasToolCalls {
+		return "tool_use"
+	}
+
+	switch reason {
+	case "stop":
+		return "end_turn"
+	case "length":
+		return "max_tokens"
+	default:
+		if reason != "" {
+			return "stop_sequence"
+		}
+		return ""
+	}
+}
+
+// StreamConverter manages state for converting Ollama streaming responses to Anthropic format
+type StreamConverter struct {
+	ID              string
+	Model           string
+	firstWrite      bool
+	contentIndex    int
+	inputTokens     int
+	outputTokens    int
+	thinkingStarted bool
+	thinkingDone    bool
+	textStarted     bool
+	toolCallsSent   map[string]bool
+}
+
+func NewStreamConverter(id, model string) *StreamConverter {
+	return &StreamConverter{
+		ID:            id,
+		Model:         model,
+		firstWrite:    true,
+		toolCallsSent: make(map[string]bool),
+	}
+}
+
+// StreamEvent represents a streaming event to be sent to the client
+type StreamEvent struct {
+	Event string
+	Data  any
+}
+
+// Process converts an Ollama ChatResponse to Anthropic streaming events
+func (c *StreamConverter) Process(r api.ChatResponse) []StreamEvent {
+	var events []StreamEvent
+
+	if c.firstWrite {
+		c.firstWrite = false
+		c.inputTokens = r.Metrics.PromptEvalCount
+
+		events = append(events, StreamEvent{
+			Event: "message_start",
+			Data: MessageStartEvent{
+				Type: "message_start",
+				Message: MessagesResponse{
+					ID:      c.ID,
+					Type:    "message",
+					Role:    "assistant",
+					Model:   c.Model,
+					Content: []ContentBlock{},
+					Usage: Usage{
+						InputTokens:  c.inputTokens,
+						OutputTokens: 0,
+					},
+				},
+			},
+		})
+	}
+
+	if r.Message.Thinking != "" && !c.thinkingDone {
+		if !c.thinkingStarted {
+			c.thinkingStarted = true
+			events = append(events, StreamEvent{
+				Event: "content_block_start",
+				Data: ContentBlockStartEvent{
+					Type:  "content_block_start",
+					Index: c.contentIndex,
+					ContentBlock: ContentBlock{
+						Type:     "thinking",
+						Thinking: ptr(""),
+					},
+				},
+			})
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type:     "thinking_delta",
+					Thinking: r.Message.Thinking,
+				},
+			},
+		})
+	}
+
+	if r.Message.Content != "" {
+		if c.thinkingStarted && !c.thinkingDone {
+			c.thinkingDone = true
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+			c.contentIndex++
+		}
+
+		if !c.textStarted {
+			c.textStarted = true
+			events = append(events, StreamEvent{
+				Event: "content_block_start",
+				Data: ContentBlockStartEvent{
+					Type:  "content_block_start",
+					Index: c.contentIndex,
+					ContentBlock: ContentBlock{
+						Type: "text",
+						Text: ptr(""),
+					},
+				},
+			})
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type: "text_delta",
+					Text: r.Message.Content,
+				},
+			},
+		})
+	}
+
+	for _, tc := range r.Message.ToolCalls {
+		if c.toolCallsSent[tc.ID] {
+			continue
+		}
+
+		if c.textStarted {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+			c.contentIndex++
+			c.textStarted = false
+		}
+
+		argsJSON, err := json.Marshal(tc.Function.Arguments)
+		if err != nil {
+			slog.Error("failed to marshal tool arguments", "error", err, "tool_id", tc.ID)
+			continue
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_start",
+			Data: ContentBlockStartEvent{
+				Type:  "content_block_start",
+				Index: c.contentIndex,
+				ContentBlock: ContentBlock{
+					Type:  "tool_use",
+					ID:    tc.ID,
+					Name:  tc.Function.Name,
+					Input: map[string]any{},
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type:        "input_json_delta",
+					PartialJSON: string(argsJSON),
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "content_block_stop",
+			Data: ContentBlockStopEvent{
+				Type:  "content_block_stop",
+				Index: c.contentIndex,
+			},
+		})
+
+		c.toolCallsSent[tc.ID] = true
+		c.contentIndex++
+	}
+
+	if r.Done {
+		if c.textStarted {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+		} else if c.thinkingStarted && !c.thinkingDone {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+		}
+
+		c.outputTokens = r.Metrics.EvalCount
+		stopReason := mapStopReason(r.DoneReason, len(c.toolCallsSent) > 0)
+
+		events = append(events, StreamEvent{
+			Event: "message_delta",
+			Data: MessageDeltaEvent{
+				Type: "message_delta",
+				Delta: MessageDelta{
+					StopReason: stopReason,
+				},
+				Usage: DeltaUsage{
+					OutputTokens: c.outputTokens,
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "message_stop",
+			Data: MessageStopEvent{
+				Type: "message_stop",
+			},
+		})
+	}
+
+	return events
+}
+
+// generateID generates a unique ID with the given prefix using crypto/rand
+func generateID(prefix string) string {
+	b := make([]byte, 12)
+	if _, err := rand.Read(b); err != nil {
+		// Fallback to time-based ID if crypto/rand fails
+		return fmt.Sprintf("%s_%d", prefix, time.Now().UnixNano())
+	}
+	return fmt.Sprintf("%s_%x", prefix, b)
+}
+
+// GenerateMessageID generates a unique message ID
+func GenerateMessageID() string {
+	return generateID("msg")
+}
+
+// ptr returns a pointer to the given string value
+func ptr(s string) *string {
+	return &s
+}
+
+// mapToArgs converts a map to ToolCallFunctionArguments
+func mapToArgs(m map[string]any) api.ToolCallFunctionArguments {
+	args := api.NewToolCallFunctionArguments()
+	for k, v := range m {
+		args.Set(k, v)
+	}
+	return args
+}

anthropic/anthropic_test.go

@@ -0,0 +1,953 @@
+package anthropic
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+const (
+	testImage = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
+)
+
+// testArgs creates ToolCallFunctionArguments from a map (convenience function for tests)
+func testArgs(m map[string]any) api.ToolCallFunctionArguments {
+	args := api.NewToolCallFunctionArguments()
+	for k, v := range m {
+		args.Set(k, v)
+	}
+	return args
+}
+
+func TestFromMessagesRequest_Basic(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Model != "test-model" {
+		t.Errorf("expected model 'test-model', got %q", result.Model)
+	}
+
+	if len(result.Messages) != 1 {
+		t.Fatalf("expected 1 message, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Role != "user" || result.Messages[0].Content != "Hello" {
+		t.Errorf("unexpected message: %+v", result.Messages[0])
+	}
+
+	if numPredict, ok := result.Options["num_predict"].(int); !ok || numPredict != 1024 {
+		t.Errorf("expected num_predict 1024, got %v", result.Options["num_predict"])
+	}
+}
+
+func TestFromMessagesRequest_WithSystemPrompt(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		System:    "You are a helpful assistant.",
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Role != "system" || result.Messages[0].Content != "You are a helpful assistant." {
+		t.Errorf("unexpected system message: %+v", result.Messages[0])
+	}
+}
+
+func TestFromMessagesRequest_WithSystemPromptArray(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		System: []any{
+			map[string]any{"type": "text", "text": "You are helpful."},
+			map[string]any{"type": "text", "text": " Be concise."},
+		},
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Content != "You are helpful. Be concise." {
+		t.Errorf("unexpected system message content: %q", result.Messages[0].Content)
+	}
+}
+
+func TestFromMessagesRequest_WithOptions(t *testing.T) {
+	temp := 0.7
+	topP := 0.9
+	topK := 40
+	req := MessagesRequest{
+		Model:         "test-model",
+		MaxTokens:     2048,
+		Messages:      []MessageParam{{Role: "user", Content: "Hello"}},
+		Temperature:   &temp,
+		TopP:          &topP,
+		TopK:          &topK,
+		StopSequences: []string{"\n", "END"},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Options["temperature"] != 0.7 {
+		t.Errorf("expected temperature 0.7, got %v", result.Options["temperature"])
+	}
+	if result.Options["top_p"] != 0.9 {
+		t.Errorf("expected top_p 0.9, got %v", result.Options["top_p"])
+	}
+	if result.Options["top_k"] != 40 {
+		t.Errorf("expected top_k 40, got %v", result.Options["top_k"])
+	}
+	if diff := cmp.Diff([]string{"\n", "END"}, result.Options["stop"]); diff != "" {
+		t.Errorf("stop sequences mismatch: %s", diff)
+	}
+}
+
+func TestFromMessagesRequest_WithImage(t *testing.T) {
+	imgData, _ := base64.StdEncoding.DecodeString(testImage)
+
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "user",
+				Content: []any{
+					map[string]any{"type": "text", "text": "What's in this 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))
+	}
+
+	if result.Messages[0].Content != "What's in this image?" {
+		t.Errorf("expected content 'What's in this image?', got %q", result.Messages[0].Content)
+	}
+
+	if len(result.Messages[0].Images) != 1 {
+		t.Fatalf("expected 1 image, got %d", len(result.Messages[0].Images))
+	}
+
+	if string(result.Messages[0].Images[0]) != string(imgData) {
+		t.Error("image data mismatch")
+	}
+}
+
+func TestFromMessagesRequest_WithToolUse(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "What's the weather in Paris?"},
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type":  "tool_use",
+						"id":    "call_123",
+						"name":  "get_weather",
+						"input": map[string]any{"location": "Paris"},
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if len(result.Messages[1].ToolCalls) != 1 {
+		t.Fatalf("expected 1 tool call, got %d", len(result.Messages[1].ToolCalls))
+	}
+
+	tc := result.Messages[1].ToolCalls[0]
+	if tc.ID != "call_123" {
+		t.Errorf("expected tool call ID 'call_123', got %q", tc.ID)
+	}
+	if tc.Function.Name != "get_weather" {
+		t.Errorf("expected tool name 'get_weather', got %q", tc.Function.Name)
+	}
+}
+
+func TestFromMessagesRequest_WithToolResult(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "user",
+				Content: []any{
+					map[string]any{
+						"type":        "tool_result",
+						"tool_use_id": "call_123",
+						"content":     "The weather in Paris is sunny, 22°C",
+					},
+				},
+			},
+		},
+	}
+
+	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_123" {
+		t.Errorf("expected tool_call_id 'call_123', got %q", msg.ToolCallID)
+	}
+	if msg.Content != "The weather in Paris is sunny, 22°C" {
+		t.Errorf("unexpected content: %q", msg.Content)
+	}
+}
+
+func TestFromMessagesRequest_WithTools(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Tools: []Tool{
+			{
+				Name:        "get_weather",
+				Description: "Get current weather",
+				InputSchema: json.RawMessage(`{"type":"object","properties":{"location":{"type":"string"}},"required":["location"]}`),
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Tools) != 1 {
+		t.Fatalf("expected 1 tool, got %d", len(result.Tools))
+	}
+
+	tool := result.Tools[0]
+	if tool.Type != "function" {
+		t.Errorf("expected type 'function', got %q", tool.Type)
+	}
+	if tool.Function.Name != "get_weather" {
+		t.Errorf("expected name 'get_weather', got %q", tool.Function.Name)
+	}
+	if tool.Function.Description != "Get current weather" {
+		t.Errorf("expected description 'Get current weather', got %q", tool.Function.Description)
+	}
+}
+
+func TestFromMessagesRequest_WithThinking(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Thinking:  &ThinkingConfig{Type: "enabled", BudgetTokens: 1000},
+	}
+
+	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 v, ok := result.Think.Value.(bool); !ok || !v {
+		t.Errorf("expected Think.Value to be true, got %v", result.Think.Value)
+	}
+}
+
+// TestFromMessagesRequest_ThinkingOnlyBlock verifies that messages containing only
+// a thinking block (no text, images, or tool calls) are preserved and not dropped.
+func TestFromMessagesRequest_ThinkingOnlyBlock(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type":     "thinking",
+						"thinking": "Let me think about this...",
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	assistantMsg := result.Messages[1]
+	if assistantMsg.Thinking != "Let me think about this..." {
+		t.Errorf("expected thinking content, got %q", assistantMsg.Thinking)
+	}
+}
+
+func TestFromMessagesRequest_ToolUseMissingID(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type": "tool_use",
+						"name": "get_weather",
+					},
+				},
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for missing tool_use id")
+	}
+	if err.Error() != "tool_use block missing required 'id' field" {
+		t.Errorf("unexpected error message: %v", err)
+	}
+}
+
+func TestFromMessagesRequest_ToolUseMissingName(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type": "tool_use",
+						"id":   "call_123",
+					},
+				},
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for missing tool_use name")
+	}
+	if err.Error() != "tool_use block missing required 'name' field" {
+		t.Errorf("unexpected error message: %v", err)
+	}
+}
+
+func TestFromMessagesRequest_InvalidToolSchema(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Tools: []Tool{
+			{
+				Name:        "bad_tool",
+				InputSchema: json.RawMessage(`{invalid json`),
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for invalid tool schema")
+	}
+}
+
+func TestToMessagesResponse_Basic(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: "Hello there!",
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics: api.Metrics{
+			PromptEvalCount: 10,
+			EvalCount:       5,
+		},
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if result.ID != "msg_123" {
+		t.Errorf("expected ID 'msg_123', got %q", result.ID)
+	}
+	if result.Type != "message" {
+		t.Errorf("expected type 'message', got %q", result.Type)
+	}
+	if result.Role != "assistant" {
+		t.Errorf("expected role 'assistant', got %q", result.Role)
+	}
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "text" || result.Content[0].Text == nil || *result.Content[0].Text != "Hello there!" {
+		t.Errorf("unexpected content: %+v", result.Content[0])
+	}
+	if result.StopReason != "end_turn" {
+		t.Errorf("expected stop_reason 'end_turn', got %q", result.StopReason)
+	}
+	if result.Usage.InputTokens != 10 || result.Usage.OutputTokens != 5 {
+		t.Errorf("unexpected usage: %+v", result.Usage)
+	}
+}
+
+func TestToMessagesResponse_WithToolCalls(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_123",
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"location": "Paris"}),
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "tool_use" {
+		t.Errorf("expected type 'tool_use', got %q", result.Content[0].Type)
+	}
+	if result.Content[0].ID != "call_123" {
+		t.Errorf("expected ID 'call_123', got %q", result.Content[0].ID)
+	}
+	if result.Content[0].Name != "get_weather" {
+		t.Errorf("expected name 'get_weather', got %q", result.Content[0].Name)
+	}
+	if result.StopReason != "tool_use" {
+		t.Errorf("expected stop_reason 'tool_use', got %q", result.StopReason)
+	}
+}
+
+func TestToMessagesResponse_WithThinking(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:     "assistant",
+			Content:  "The answer is 42.",
+			Thinking: "Let me think about this...",
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if len(result.Content) != 2 {
+		t.Fatalf("expected 2 content blocks, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "thinking" {
+		t.Errorf("expected first block type 'thinking', got %q", result.Content[0].Type)
+	}
+	if result.Content[0].Thinking == nil || *result.Content[0].Thinking != "Let me think about this..." {
+		t.Errorf("unexpected thinking content: %v", result.Content[0].Thinking)
+	}
+	if result.Content[1].Type != "text" {
+		t.Errorf("expected second block type 'text', got %q", result.Content[1].Type)
+	}
+}
+
+func TestMapStopReason(t *testing.T) {
+	tests := []struct {
+		reason       string
+		hasToolCalls bool
+		want         string
+	}{
+		{"stop", false, "end_turn"},
+		{"length", false, "max_tokens"},
+		{"stop", true, "tool_use"},
+		{"other", false, "stop_sequence"},
+		{"", false, ""},
+	}
+
+	for _, tt := range tests {
+		got := mapStopReason(tt.reason, tt.hasToolCalls)
+		if got != tt.want {
+			t.Errorf("mapStopReason(%q, %v) = %q, want %q", tt.reason, tt.hasToolCalls, got, tt.want)
+		}
+	}
+}
+
+func TestNewError(t *testing.T) {
+	tests := []struct {
+		code int
+		want string
+	}{
+		{400, "invalid_request_error"},
+		{401, "authentication_error"},
+		{403, "permission_error"},
+		{404, "not_found_error"},
+		{429, "rate_limit_error"},
+		{500, "api_error"},
+		{503, "overloaded_error"},
+		{529, "overloaded_error"},
+	}
+
+	for _, tt := range tests {
+		result := NewError(tt.code, "test message")
+		if result.Type != "error" {
+			t.Errorf("NewError(%d) type = %q, want 'error'", tt.code, result.Type)
+		}
+		if result.Error.Type != tt.want {
+			t.Errorf("NewError(%d) error.type = %q, want %q", tt.code, result.Error.Type, tt.want)
+		}
+		if result.Error.Message != "test message" {
+			t.Errorf("NewError(%d) message = %q, want 'test message'", tt.code, result.Error.Message)
+		}
+		if result.RequestID == "" {
+			t.Errorf("NewError(%d) request_id should not be empty", tt.code)
+		}
+	}
+}
+
+func TestGenerateMessageID(t *testing.T) {
+	id1 := GenerateMessageID()
+	id2 := GenerateMessageID()
+
+	if id1 == "" {
+		t.Error("GenerateMessageID returned empty string")
+	}
+	if id1 == id2 {
+		t.Error("GenerateMessageID returned duplicate IDs")
+	}
+	if len(id1) < 10 {
+		t.Errorf("GenerateMessageID returned short ID: %q", id1)
+	}
+	if id1[:4] != "msg_" {
+		t.Errorf("GenerateMessageID should start with 'msg_', got %q", id1[:4])
+	}
+}
+
+func TestStreamConverter_Basic(t *testing.T) {
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	// First chunk
+	resp1 := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: "Hello",
+		},
+		Metrics: api.Metrics{PromptEvalCount: 10},
+	}
+
+	events1 := conv.Process(resp1)
+	if len(events1) < 3 {
+		t.Fatalf("expected at least 3 events for first chunk, got %d", len(events1))
+	}
+
+	// Should have message_start, content_block_start, content_block_delta
+	if events1[0].Event != "message_start" {
+		t.Errorf("expected first event 'message_start', got %q", events1[0].Event)
+	}
+	if events1[1].Event != "content_block_start" {
+		t.Errorf("expected second event 'content_block_start', got %q", events1[1].Event)
+	}
+	if events1[2].Event != "content_block_delta" {
+		t.Errorf("expected third event 'content_block_delta', got %q", events1[2].Event)
+	}
+
+	// Final chunk
+	resp2 := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: " world!",
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics:    api.Metrics{EvalCount: 5},
+	}
+
+	events2 := conv.Process(resp2)
+
+	// Should have content_block_delta, content_block_stop, message_delta, message_stop
+	hasStop := false
+	for _, e := range events2 {
+		if e.Event == "message_stop" {
+			hasStop = true
+		}
+	}
+	if !hasStop {
+		t.Error("expected message_stop event in final chunk")
+	}
+}
+
+func TestStreamConverter_WithToolCalls(t *testing.T) {
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_123",
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"location": "Paris"}),
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics:    api.Metrics{PromptEvalCount: 10, EvalCount: 5},
+	}
+
+	events := conv.Process(resp)
+
+	hasToolStart := false
+	hasToolDelta := false
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					hasToolStart = true
+				}
+			}
+		}
+		if e.Event == "content_block_delta" {
+			if delta, ok := e.Data.(ContentBlockDeltaEvent); ok {
+				if delta.Delta.Type == "input_json_delta" {
+					hasToolDelta = true
+				}
+			}
+		}
+	}
+
+	if !hasToolStart {
+		t.Error("expected tool_use content_block_start event")
+	}
+	if !hasToolDelta {
+		t.Error("expected input_json_delta event")
+	}
+}
+
+func TestStreamConverter_ToolCallWithUnmarshalableArgs(t *testing.T) {
+	// Test that unmarshalable arguments (like channels) are handled gracefully
+	// and don't cause a panic or corrupt stream
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	// Create a channel which cannot be JSON marshaled
+	unmarshalable := make(chan int)
+	badArgs := api.NewToolCallFunctionArguments()
+	badArgs.Set("channel", unmarshalable)
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_bad",
+					Function: api.ToolCallFunction{
+						Name:      "bad_function",
+						Arguments: badArgs,
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	// Should not panic and should skip the unmarshalable tool call
+	events := conv.Process(resp)
+
+	// Verify no tool_use block was started (since marshal failed before block start)
+	hasToolStart := false
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					hasToolStart = true
+				}
+			}
+		}
+	}
+
+	if hasToolStart {
+		t.Error("expected no tool_use block when arguments cannot be marshaled")
+	}
+}
+
+func TestStreamConverter_MultipleToolCallsWithMixedValidity(t *testing.T) {
+	// Test that valid tool calls still work when mixed with invalid ones
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	unmarshalable := make(chan int)
+	badArgs := api.NewToolCallFunctionArguments()
+	badArgs.Set("channel", unmarshalable)
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_good",
+					Function: api.ToolCallFunction{
+						Name:      "good_function",
+						Arguments: testArgs(map[string]any{"location": "Paris"}),
+					},
+				},
+				{
+					ID: "call_bad",
+					Function: api.ToolCallFunction{
+						Name:      "bad_function",
+						Arguments: badArgs,
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	events := conv.Process(resp)
+
+	// Count tool_use blocks - should only have 1 (the valid one)
+	toolStartCount := 0
+	toolDeltaCount := 0
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					toolStartCount++
+					if start.ContentBlock.Name != "good_function" {
+						t.Errorf("expected tool name 'good_function', got %q", start.ContentBlock.Name)
+					}
+				}
+			}
+		}
+		if e.Event == "content_block_delta" {
+			if delta, ok := e.Data.(ContentBlockDeltaEvent); ok {
+				if delta.Delta.Type == "input_json_delta" {
+					toolDeltaCount++
+				}
+			}
+		}
+	}
+
+	if toolStartCount != 1 {
+		t.Errorf("expected 1 tool_use block, got %d", toolStartCount)
+	}
+	if toolDeltaCount != 1 {
+		t.Errorf("expected 1 input_json_delta, got %d", toolDeltaCount)
+	}
+}
+
+// TestContentBlockJSON_EmptyFieldsPresent verifies that empty text and thinking fields
+// are serialized in JSON output. The Anthropic SDK requires these fields to be present
+// (even when empty) in content_block_start events to properly accumulate streaming deltas.
+// Without these fields, the SDK throws: "TypeError: unsupported operand type(s) for +=: 'NoneType' and 'str'"
+func TestContentBlockJSON_EmptyFieldsPresent(t *testing.T) {
+	tests := []struct {
+		name     string
+		block    ContentBlock
+		wantKeys []string
+	}{
+		{
+			name: "text block includes empty text field",
+			block: ContentBlock{
+				Type: "text",
+				Text: ptr(""),
+			},
+			wantKeys: []string{"type", "text"},
+		},
+		{
+			name: "thinking block includes empty thinking field",
+			block: ContentBlock{
+				Type:     "thinking",
+				Thinking: ptr(""),
+			},
+			wantKeys: []string{"type", "thinking"},
+		},
+		{
+			name: "text block with content",
+			block: ContentBlock{
+				Type: "text",
+				Text: ptr("hello"),
+			},
+			wantKeys: []string{"type", "text"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			data, err := json.Marshal(tt.block)
+			if err != nil {
+				t.Fatalf("failed to marshal: %v", err)
+			}
+
+			var result map[string]any
+			if err := json.Unmarshal(data, &result); err != nil {
+				t.Fatalf("failed to unmarshal: %v", err)
+			}
+
+			for _, key := range tt.wantKeys {
+				if _, ok := result[key]; !ok {
+					t.Errorf("expected key %q to be present in JSON output, got: %s", key, string(data))
+				}
+			}
+		})
+	}
+}
+
+// TestStreamConverter_ContentBlockStartIncludesEmptyFields verifies that content_block_start
+// events include the required empty fields for SDK compatibility.
+func TestStreamConverter_ContentBlockStartIncludesEmptyFields(t *testing.T) {
+	t.Run("text block start includes empty text", func(t *testing.T) {
+		conv := NewStreamConverter("msg_123", "test-model")
+
+		resp := api.ChatResponse{
+			Model:   "test-model",
+			Message: api.Message{Role: "assistant", Content: "hello"},
+		}
+
+		events := conv.Process(resp)
+
+		var foundTextStart bool
+		for _, e := range events {
+			if e.Event == "content_block_start" {
+				if start, ok := e.Data.(ContentBlockStartEvent); ok {
+					if start.ContentBlock.Type == "text" {
+						foundTextStart = true
+						// Marshal and verify the text field is present
+						data, _ := json.Marshal(start)
+						var result map[string]any
+						json.Unmarshal(data, &result)
+						cb := result["content_block"].(map[string]any)
+						if _, ok := cb["text"]; !ok {
+							t.Error("content_block_start for text should include 'text' field")
+						}
+					}
+				}
+			}
+		}
+
+		if !foundTextStart {
+			t.Error("expected text content_block_start event")
+		}
+	})
+
+	t.Run("thinking block start includes empty thinking", func(t *testing.T) {
+		conv := NewStreamConverter("msg_123", "test-model")
+
+		resp := api.ChatResponse{
+			Model:   "test-model",
+			Message: api.Message{Role: "assistant", Thinking: "let me think..."},
+		}
+
+		events := conv.Process(resp)
+
+		var foundThinkingStart bool
+		for _, e := range events {
+			if e.Event == "content_block_start" {
+				if start, ok := e.Data.(ContentBlockStartEvent); ok {
+					if start.ContentBlock.Type == "thinking" {
+						foundThinkingStart = true
+						data, _ := json.Marshal(start)
+						var result map[string]any
+						json.Unmarshal(data, &result)
+						cb := result["content_block"].(map[string]any)
+						if _, ok := cb["thinking"]; !ok {
+							t.Error("content_block_start for thinking should include 'thinking' field")
+						}
+					}
+				}
+			}
+		}
+
+		if !foundThinkingStart {
+			t.Error("expected thinking content_block_start event")
+		}
+	})
+}

api/client.go

@@ -165,7 +165,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	return nil
 }
 
-const maxBufferSize = 512 * format.KiloByte
+const maxBufferSize = 8 * format.MegaByte
 
 func (c *Client) stream(ctx context.Context, method, path string, data any, fn func([]byte) error) error {
 	var buf io.Reader

api/types.go

@@ -127,6 +127,20 @@ type GenerateRequest struct {
 	// each with an associated log probability. Only applies when Logprobs is true.
 	// Valid values are 0-20. Default is 0 (only return the selected token's logprob).
 	TopLogprobs int `json:"top_logprobs,omitempty"`
+
+	// Experimental: Image generation fields (may change or be removed)
+
+	// Width is the width of the generated image in pixels.
+	// Only used for image generation models.
+	Width int32 `json:"width,omitempty"`
+
+	// Height is the height of the generated image in pixels.
+	// Only used for image generation models.
+	Height int32 `json:"height,omitempty"`
+
+	// Steps is the number of diffusion steps for image generation.
+	// Only used for image generation models.
+	Steps int32 `json:"steps,omitempty"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -735,7 +749,7 @@ type ShowResponse struct {
 	Messages      []Message          `json:"messages,omitempty"`
 	RemoteModel   string             `json:"remote_model,omitempty"`
 	RemoteHost    string             `json:"remote_host,omitempty"`
-	ModelInfo     map[string]any     `json:"model_info,omitempty"`
+	ModelInfo     map[string]any     `json:"model_info"`
 	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
 	Tensors       []Tensor           `json:"tensors,omitempty"`
 	Capabilities  []model.Capability `json:"capabilities,omitempty"`
@@ -860,6 +874,20 @@ type GenerateResponse struct {
 	// Logprobs contains log probability information for the generated tokens,
 	// if requested via the Logprobs parameter.
 	Logprobs []Logprob `json:"logprobs,omitempty"`
+
+	// Experimental: Image generation fields (may change or be removed)
+
+	// Image contains a base64-encoded generated image.
+	// Only present for image generation models.
+	Image string `json:"image,omitempty"`
+
+	// Completed is the number of completed steps in image generation.
+	// Only present for image generation models during streaming.
+	Completed int64 `json:"completed,omitempty"`
+
+	// Total is the total number of steps for image generation.
+	// Only present for image generation models during streaming.
+	Total int64 `json:"total,omitempty"`
 }
 
 // ModelDetails provides details about a model.

app/cmd/app/app_darwin.m

@@ -14,6 +14,7 @@ extern NSString *SystemWidePath;
 @interface AppDelegate () <NSWindowDelegate, WKNavigationDelegate, WKUIDelegate>
 @property(strong, nonatomic) NSStatusItem *statusItem;
 @property(assign, nonatomic) BOOL updateAvailable;
+@property(assign, nonatomic) BOOL systemShutdownInProgress;
 @end
 
 @implementation AppDelegate
@@ -40,6 +41,13 @@ bool firstTimeRun,startHidden; // Set in run before initialization
 }
 
 - (void)applicationDidFinishLaunching:(NSNotification *)aNotification {
+    // Register for system shutdown/restart notification so we can allow termination
+    [[[NSWorkspace sharedWorkspace] notificationCenter]
+        addObserver:self
+           selector:@selector(systemWillPowerOff:)
+               name:NSWorkspaceWillPowerOffNotification
+             object:nil];
+
     // if we're in development mode, set the app icon
     NSString *bundlePath = [[NSBundle mainBundle] bundlePath];
     if (![bundlePath hasSuffix:@".app"]) {
@@ -278,7 +286,18 @@ bool firstTimeRun,startHidden; // Set in run before initialization
     [NSApp activateIgnoringOtherApps:YES];
 }
 
+- (void)systemWillPowerOff:(NSNotification *)notification {
+    // Set flag so applicationShouldTerminate: knows to allow termination.
+    // The system will call applicationShouldTerminate: after posting this notification.
+    self.systemShutdownInProgress = YES;
+}
+
 - (NSApplicationTerminateReply)applicationShouldTerminate:(NSApplication *)sender {
+    // Allow termination if the system is shutting down or restarting
+    if (self.systemShutdownInProgress) {
+        return NSTerminateNow;
+    }
+    // Otherwise just hide the app (for Cmd+Q, close button, etc.)
     [NSApp hide:nil];
     [NSApp setActivationPolicy:NSApplicationActivationPolicyAccessory];
     return NSTerminateCancel;

cmd/cmd.go

@@ -35,6 +35,7 @@ import (
 	"golang.org/x/term"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/cmd/config"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/parser"
@@ -46,6 +47,9 @@ import (
 	"github.com/ollama/ollama/types/syncmap"
 	"github.com/ollama/ollama/version"
 	xcmd "github.com/ollama/ollama/x/cmd"
+	"github.com/ollama/ollama/x/create"
+	xcreateclient "github.com/ollama/ollama/x/create/client"
+	"github.com/ollama/ollama/x/imagegen"
 )
 
 const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
@@ -91,11 +95,88 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
+	// Validate model name early to fail fast
+	modelName := args[0]
+	name := model.ParseName(modelName)
+	if !name.IsValid() {
+		return fmt.Errorf("invalid model name: %s", modelName)
+	}
+
+	// Check for --experimental flag for safetensors model creation
+	experimental, _ := cmd.Flags().GetBool("experimental")
+	if experimental {
+		// Get Modelfile content - either from -f flag or default to "FROM ."
+		var reader io.Reader
+		filename, err := getModelfileName(cmd)
+		if os.IsNotExist(err) || filename == "" {
+			// No Modelfile specified or found - use default
+			reader = strings.NewReader("FROM .\n")
+		} else if err != nil {
+			return err
+		} else {
+			f, err := os.Open(filename)
+			if err != nil {
+				return err
+			}
+			defer f.Close()
+			reader = f
+		}
+
+		// Parse the Modelfile
+		modelfile, err := parser.ParseFile(reader)
+		if err != nil {
+			return fmt.Errorf("failed to parse Modelfile: %w", err)
+		}
+
+		// Extract FROM path and configuration
+		var modelDir string
+		mfConfig := &xcreateclient.ModelfileConfig{}
+
+		for _, cmd := range modelfile.Commands {
+			switch cmd.Name {
+			case "model":
+				modelDir = cmd.Args
+			case "template":
+				mfConfig.Template = cmd.Args
+			case "system":
+				mfConfig.System = cmd.Args
+			case "license":
+				mfConfig.License = cmd.Args
+			}
+		}
+
+		if modelDir == "" {
+			modelDir = "."
+		}
+
+		// Resolve relative paths based on Modelfile location
+		if !filepath.IsAbs(modelDir) && filename != "" {
+			modelDir = filepath.Join(filepath.Dir(filename), modelDir)
+		}
+
+		quantize, _ := cmd.Flags().GetString("quantize")
+		return xcreateclient.CreateModel(xcreateclient.CreateOptions{
+			ModelName: modelName,
+			ModelDir:  modelDir,
+			Quantize:  quantize,
+			Modelfile: mfConfig,
+		}, p)
+	}
+
 	var reader io.Reader
 
 	filename, err := getModelfileName(cmd)
 	if os.IsNotExist(err) {
 		if filename == "" {
+			// No Modelfile found - check if current directory is an image gen model
+			if create.IsTensorModelDir(".") {
+				quantize, _ := cmd.Flags().GetString("quantize")
+				return xcreateclient.CreateModel(xcreateclient.CreateOptions{
+					ModelName: modelName,
+					ModelDir:  ".",
+					Quantize:  quantize,
+				}, p)
+			}
 			reader = strings.NewReader("FROM .\n")
 		} else {
 			return errModelfileNotFound
@@ -127,7 +208,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	}
 	spinner.Stop()
 
-	req.Model = args[0]
+	req.Model = modelName
 	quantize, _ := cmd.Flags().GetString("quantize")
 	if quantize != "" {
 		req.Quantize = quantize
@@ -457,6 +538,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	name := args[0]
+
 	info, err := func() (*api.ShowResponse, error) {
 		showReq := &api.ShowRequest{Name: name}
 		info, err := client.Show(cmd.Context(), showReq)
@@ -518,9 +600,18 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return generateEmbedding(cmd, name, opts.Prompt, opts.KeepAlive, truncate, dimensions)
 	}
 
+	// Check if this is an image generation model
+	if slices.Contains(info.Capabilities, model.CapabilityImage) {
+		if opts.Prompt == "" && !interactive {
+			return errors.New("image generation models require a prompt. Usage: ollama run " + name + " \"your prompt here\"")
+		}
+		return imagegen.RunCLI(cmd, name, opts.Prompt, interactive, opts.KeepAlive)
+	}
+
 	// Check for experimental flag
 	isExperimental, _ := cmd.Flags().GetBool("experimental")
-	yoloMode, _ := cmd.Flags().GetBool("yolo")
+	yoloMode, _ := cmd.Flags().GetBool("experimental-yolo")
+	enableWebsearch, _ := cmd.Flags().GetBool("experimental-websearch")
 
 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
@@ -550,7 +641,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 
 		// Use experimental agent loop with tools
 		if isExperimental {
-			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive, yoloMode)
+			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive, yoloMode, enableWebsearch)
 		}
 
 		return generateInteractive(cmd, opts)
@@ -656,7 +747,11 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 
 			bar, ok := bars[resp.Digest]
 			if !ok {
-				bar = progress.NewBar(fmt.Sprintf("pushing %s...", resp.Digest[7:19]), resp.Total, resp.Completed)
+				msg := resp.Status
+				if msg == "" {
+					msg = fmt.Sprintf("pushing %s...", resp.Digest[7:19])
+				}
+				bar = progress.NewBar(msg, resp.Total, resp.Completed)
 				bars[resp.Digest] = bar
 				p.Add(resp.Digest, bar)
 			}
@@ -805,11 +900,11 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 	for _, arg := range args {
 		// Unload the model if it's running before deletion
 		if err := loadOrUnloadModel(cmd, &runOptions{
-			Model:     args[0],
+			Model:     arg,
 			KeepAlive: &api.Duration{Duration: 0},
 		}); err != nil {
 			if !strings.Contains(strings.ToLower(err.Error()), "not found") {
-				fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", args[0])
+				fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", arg)
 			}
 		}
 
@@ -1721,15 +1816,22 @@ func NewCLI() *cobra.Command {
 	rootCmd.Flags().BoolP("version", "v", false, "Show version information")
 
 	createCmd := &cobra.Command{
-		Use:     "create MODEL",
-		Short:   "Create a model",
-		Args:    cobra.ExactArgs(1),
-		PreRunE: checkServerHeartbeat,
-		RunE:    CreateHandler,
+		Use:   "create MODEL",
+		Short: "Create a model",
+		Args:  cobra.ExactArgs(1),
+		PreRunE: func(cmd *cobra.Command, args []string) error {
+			// Skip server check for experimental mode (writes directly to disk)
+			if experimental, _ := cmd.Flags().GetBool("experimental"); experimental {
+				return nil
+			}
+			return checkServerHeartbeat(cmd, args)
+		},
+		RunE: CreateHandler,
 	}
 
 	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\")")
 	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_K_M)")
+	createCmd.Flags().Bool("experimental", false, "Enable experimental safetensors model creation")
 
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@@ -1765,7 +1867,11 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("truncate", false, "For embedding models: truncate inputs exceeding context length (default: true). Set --truncate=false to error instead")
 	runCmd.Flags().Int("dimensions", 0, "Truncate output embeddings to specified dimension (embedding models only)")
 	runCmd.Flags().Bool("experimental", false, "Enable experimental agent loop with tools")
-	runCmd.Flags().BoolP("yolo", "y", false, "Skip all tool approval prompts (use with caution)")
+	runCmd.Flags().Bool("experimental-yolo", false, "Skip all tool approval prompts (use with caution)")
+	runCmd.Flags().Bool("experimental-websearch", false, "Enable web search tool in experimental mode")
+
+	// Image generation flags (width, height, steps, seed, etc.)
+	imagegen.RegisterFlags(runCmd)
 
 	stopCmd := &cobra.Command{
 		Use:     "stop MODEL",
@@ -1880,6 +1986,7 @@ func NewCLI() *cobra.Command {
 	} {
 		switch cmd {
 		case runCmd:
+			imagegen.AppendFlagsDocs(cmd)
 			appendEnvDocs(cmd, []envconfig.EnvVar{envVars["OLLAMA_HOST"], envVars["OLLAMA_NOHISTORY"]})
 		case serveCmd:
 			appendEnvDocs(cmd, []envconfig.EnvVar{
@@ -1920,6 +2027,7 @@ func NewCLI() *cobra.Command {
 		copyCmd,
 		deleteCmd,
 		runnerCmd,
+		config.ConfigCmd(checkServerHeartbeat),
 	)
 
 	return rootCmd

cmd/cmd_test.go

@@ -1547,6 +1547,79 @@ func TestRunOptions_Copy_ThinkValueVariants(t *testing.T) {
 	}
 }
 
+func TestShowInfoImageGen(t *testing.T) {
+	var b bytes.Buffer
+	err := showInfo(&api.ShowResponse{
+		Details: api.ModelDetails{
+			Family:            "ZImagePipeline",
+			ParameterSize:     "10.3B",
+			QuantizationLevel: "FP8",
+		},
+		Capabilities: []model.Capability{model.CapabilityImage},
+		Requires:     "0.14.0",
+	}, false, &b)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	expect := "  Model\n" +
+		"    architecture    ZImagePipeline    \n" +
+		"    parameters      10.3B             \n" +
+		"    quantization    FP8               \n" +
+		"    requires        0.14.0            \n" +
+		"\n" +
+		"  Capabilities\n" +
+		"    image    \n" +
+		"\n"
+	if diff := cmp.Diff(expect, b.String()); diff != "" {
+		t.Errorf("unexpected output (-want +got):\n%s", diff)
+	}
+}
+
+func TestPushProgressMessage(t *testing.T) {
+	tests := []struct {
+		name    string
+		status  string
+		digest  string
+		wantMsg string
+	}{
+		{
+			name:    "uses status when provided",
+			status:  "uploading model",
+			digest:  "sha256:abc123456789def",
+			wantMsg: "uploading model",
+		},
+		{
+			name:    "falls back to digest when status empty",
+			status:  "",
+			digest:  "sha256:abc123456789def",
+			wantMsg: "pushing abc123456789...",
+		},
+		{
+			name:    "handles short digest gracefully",
+			status:  "",
+			digest:  "sha256:abc",
+			wantMsg: "pushing sha256:abc...",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			msg := tt.status
+			if msg == "" {
+				if len(tt.digest) >= 19 {
+					msg = fmt.Sprintf("pushing %s...", tt.digest[7:19])
+				} else {
+					msg = fmt.Sprintf("pushing %s...", tt.digest)
+				}
+			}
+			if msg != tt.wantMsg {
+				t.Errorf("got %q, want %q", msg, tt.wantMsg)
+			}
+		})
+	}
+}
+
 func TestRunOptions_Copy_Independence(t *testing.T) {
 	// Test that modifications to original don't affect copy
 	originalThink := &api.ThinkValue{Value: "original"}

cmd/config/claude.go

@@ -0,0 +1,36 @@
+package config
+
+import (
+	"fmt"
+	"os"
+	"os/exec"
+)
+
+// Claude implements Runner for Claude Code integration
+type Claude struct{}
+
+func (c *Claude) String() string { return "Claude Code" }
+
+func (c *Claude) args(model string) []string {
+	if model != "" {
+		return []string{"--model", model}
+	}
+	return nil
+}
+
+func (c *Claude) Run(model string) error {
+	if _, err := exec.LookPath("claude"); err != nil {
+		return fmt.Errorf("claude is not installed, install from https://code.claude.com/docs/en/quickstart")
+	}
+
+	cmd := exec.Command("claude", c.args(model)...)
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	cmd.Env = append(os.Environ(),
+		"ANTHROPIC_BASE_URL=http://localhost:11434",
+		"ANTHROPIC_API_KEY=",
+		"ANTHROPIC_AUTH_TOKEN=ollama",
+	)
+	return cmd.Run()
+}

cmd/config/claude_test.go

@@ -0,0 +1,42 @@
+package config
+
+import (
+	"slices"
+	"testing"
+)
+
+func TestClaudeIntegration(t *testing.T) {
+	c := &Claude{}
+
+	t.Run("String", func(t *testing.T) {
+		if got := c.String(); got != "Claude Code" {
+			t.Errorf("String() = %q, want %q", got, "Claude Code")
+		}
+	})
+
+	t.Run("implements Runner", func(t *testing.T) {
+		var _ Runner = c
+	})
+}
+
+func TestClaudeArgs(t *testing.T) {
+	c := &Claude{}
+
+	tests := []struct {
+		name  string
+		model string
+		want  []string
+	}{
+		{"with model", "llama3.2", []string{"--model", "llama3.2"}},
+		{"empty model", "", nil},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := c.args(tt.model)
+			if !slices.Equal(got, tt.want) {
+				t.Errorf("args(%q) = %v, want %v", tt.model, got, tt.want)
+			}
+		})
+	}
+}

cmd/config/codex.go

@@ -0,0 +1,61 @@
+package config
+
+import (
+	"fmt"
+	"os"
+	"os/exec"
+	"strings"
+
+	"golang.org/x/mod/semver"
+)
+
+// Codex implements Runner for Codex integration
+type Codex struct{}
+
+func (c *Codex) String() string { return "Codex" }
+
+func (c *Codex) args(model string) []string {
+	args := []string{"--oss"}
+	if model != "" {
+		args = append(args, "-m", model)
+	}
+	return args
+}
+
+func (c *Codex) Run(model string) error {
+	if err := checkCodexVersion(); err != nil {
+		return err
+	}
+
+	cmd := exec.Command("codex", c.args(model)...)
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	return cmd.Run()
+}
+
+func checkCodexVersion() error {
+	if _, err := exec.LookPath("codex"); err != nil {
+		return fmt.Errorf("codex is not installed, install with: npm install -g @openai/codex")
+	}
+
+	out, err := exec.Command("codex", "--version").Output()
+	if err != nil {
+		return fmt.Errorf("failed to get codex version: %w", err)
+	}
+
+	// Parse output like "codex-cli 0.87.0"
+	fields := strings.Fields(strings.TrimSpace(string(out)))
+	if len(fields) < 2 {
+		return fmt.Errorf("unexpected codex version output: %s", string(out))
+	}
+
+	version := "v" + fields[len(fields)-1]
+	minVersion := "v0.81.0"
+
+	if semver.Compare(version, minVersion) < 0 {
+		return fmt.Errorf("codex version %s is too old, minimum required is %s, update with: npm update -g @openai/codex", fields[len(fields)-1], "0.81.0")
+	}
+
+	return nil
+}

cmd/config/codex_test.go

@@ -0,0 +1,28 @@
+package config
+
+import (
+	"slices"
+	"testing"
+)
+
+func TestCodexArgs(t *testing.T) {
+	c := &Codex{}
+
+	tests := []struct {
+		name  string
+		model string
+		want  []string
+	}{
+		{"with model", "llama3.2", []string{"--oss", "-m", "llama3.2"}},
+		{"empty model", "", []string{"--oss"}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := c.args(tt.model)
+			if !slices.Equal(got, tt.want) {
+				t.Errorf("args(%q) = %v, want %v", tt.model, got, tt.want)
+			}
+		})
+	}
+}

cmd/config/config.go

@@ -0,0 +1,115 @@
+// Package config provides integration configuration for external coding tools
+// (Claude Code, Codex, Droid, OpenCode) to use Ollama models.
+package config
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+type integration struct {
+	Models []string `json:"models"`
+}
+
+type config struct {
+	Integrations map[string]*integration `json:"integrations"`
+}
+
+func configPath() (string, error) {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(home, ".ollama", "config", "config.json"), nil
+}
+
+func load() (*config, error) {
+	path, err := configPath()
+	if err != nil {
+		return nil, err
+	}
+
+	data, err := os.ReadFile(path)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return &config{Integrations: make(map[string]*integration)}, nil
+		}
+		return nil, err
+	}
+
+	var cfg config
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("failed to parse config: %w, at: %s", err, path)
+	}
+	if cfg.Integrations == nil {
+		cfg.Integrations = make(map[string]*integration)
+	}
+	return &cfg, nil
+}
+
+func save(cfg *config) error {
+	path, err := configPath()
+	if err != nil {
+		return err
+	}
+
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		return err
+	}
+
+	data, err := json.MarshalIndent(cfg, "", "  ")
+	if err != nil {
+		return err
+	}
+
+	return writeWithBackup(path, data)
+}
+
+func saveIntegration(appName string, models []string) error {
+	if appName == "" {
+		return errors.New("app name cannot be empty")
+	}
+
+	cfg, err := load()
+	if err != nil {
+		return err
+	}
+
+	cfg.Integrations[strings.ToLower(appName)] = &integration{
+		Models: models,
+	}
+
+	return save(cfg)
+}
+
+func loadIntegration(appName string) (*integration, error) {
+	cfg, err := load()
+	if err != nil {
+		return nil, err
+	}
+
+	ic, ok := cfg.Integrations[strings.ToLower(appName)]
+	if !ok {
+		return nil, os.ErrNotExist
+	}
+
+	return ic, nil
+}
+
+func listIntegrations() ([]integration, error) {
+	cfg, err := load()
+	if err != nil {
+		return nil, err
+	}
+
+	result := make([]integration, 0, len(cfg.Integrations))
+	for _, ic := range cfg.Integrations {
+		result = append(result, *ic)
+	}
+
+	return result, nil
+}

cmd/config/config_test.go

@@ -0,0 +1,373 @@
+package config
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// setTestHome sets both HOME (Unix) and USERPROFILE (Windows) for cross-platform tests
+func setTestHome(t *testing.T, dir string) {
+	t.Setenv("HOME", dir)
+	t.Setenv("USERPROFILE", dir)
+}
+
+// editorPaths is a test helper that safely calls Paths if the runner implements Editor
+func editorPaths(r Runner) []string {
+	if editor, ok := r.(Editor); ok {
+		return editor.Paths()
+	}
+	return nil
+}
+
+func TestIntegrationConfig(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("save and load round-trip", func(t *testing.T) {
+		models := []string{"llama3.2", "mistral", "qwen2.5"}
+		if err := saveIntegration("claude", models); err != nil {
+			t.Fatal(err)
+		}
+
+		config, err := loadIntegration("claude")
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if len(config.Models) != len(models) {
+			t.Errorf("expected %d models, got %d", len(models), len(config.Models))
+		}
+		for i, m := range models {
+			if config.Models[i] != m {
+				t.Errorf("model %d: expected %s, got %s", i, m, config.Models[i])
+			}
+		}
+	})
+
+	t.Run("defaultModel returns first model", func(t *testing.T) {
+		saveIntegration("codex", []string{"model-a", "model-b"})
+
+		config, _ := loadIntegration("codex")
+		defaultModel := ""
+		if len(config.Models) > 0 {
+			defaultModel = config.Models[0]
+		}
+		if defaultModel != "model-a" {
+			t.Errorf("expected model-a, got %s", defaultModel)
+		}
+	})
+
+	t.Run("defaultModel returns empty for no models", func(t *testing.T) {
+		config := &integration{Models: []string{}}
+		defaultModel := ""
+		if len(config.Models) > 0 {
+			defaultModel = config.Models[0]
+		}
+		if defaultModel != "" {
+			t.Errorf("expected empty string, got %s", defaultModel)
+		}
+	})
+
+	t.Run("app name is case-insensitive", func(t *testing.T) {
+		saveIntegration("Claude", []string{"model-x"})
+
+		config, err := loadIntegration("claude")
+		if err != nil {
+			t.Fatal(err)
+		}
+		defaultModel := ""
+		if len(config.Models) > 0 {
+			defaultModel = config.Models[0]
+		}
+		if defaultModel != "model-x" {
+			t.Errorf("expected model-x, got %s", defaultModel)
+		}
+	})
+
+	t.Run("multiple integrations in single file", func(t *testing.T) {
+		saveIntegration("app1", []string{"model-1"})
+		saveIntegration("app2", []string{"model-2"})
+
+		config1, _ := loadIntegration("app1")
+		config2, _ := loadIntegration("app2")
+
+		defaultModel1 := ""
+		if len(config1.Models) > 0 {
+			defaultModel1 = config1.Models[0]
+		}
+		defaultModel2 := ""
+		if len(config2.Models) > 0 {
+			defaultModel2 = config2.Models[0]
+		}
+		if defaultModel1 != "model-1" {
+			t.Errorf("expected model-1, got %s", defaultModel1)
+		}
+		if defaultModel2 != "model-2" {
+			t.Errorf("expected model-2, got %s", defaultModel2)
+		}
+	})
+}
+
+func TestListIntegrations(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("returns empty when no integrations", func(t *testing.T) {
+		configs, err := listIntegrations()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(configs) != 0 {
+			t.Errorf("expected 0 integrations, got %d", len(configs))
+		}
+	})
+
+	t.Run("returns all saved integrations", func(t *testing.T) {
+		saveIntegration("claude", []string{"model-1"})
+		saveIntegration("droid", []string{"model-2"})
+
+		configs, err := listIntegrations()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(configs) != 2 {
+			t.Errorf("expected 2 integrations, got %d", len(configs))
+		}
+	})
+}
+
+func TestEditorPaths(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("returns empty for claude (no Editor)", func(t *testing.T) {
+		r := integrations["claude"]
+		paths := editorPaths(r)
+		if len(paths) != 0 {
+			t.Errorf("expected no paths for claude, got %v", paths)
+		}
+	})
+
+	t.Run("returns empty for codex (no Editor)", func(t *testing.T) {
+		r := integrations["codex"]
+		paths := editorPaths(r)
+		if len(paths) != 0 {
+			t.Errorf("expected no paths for codex, got %v", paths)
+		}
+	})
+
+	t.Run("returns empty for droid when no config exists", func(t *testing.T) {
+		r := integrations["droid"]
+		paths := editorPaths(r)
+		if len(paths) != 0 {
+			t.Errorf("expected no paths, got %v", paths)
+		}
+	})
+
+	t.Run("returns path for droid when config exists", func(t *testing.T) {
+		settingsDir, _ := os.UserHomeDir()
+		settingsDir = filepath.Join(settingsDir, ".factory")
+		os.MkdirAll(settingsDir, 0o755)
+		os.WriteFile(filepath.Join(settingsDir, "settings.json"), []byte(`{}`), 0o644)
+
+		r := integrations["droid"]
+		paths := editorPaths(r)
+		if len(paths) != 1 {
+			t.Errorf("expected 1 path, got %d", len(paths))
+		}
+	})
+
+	t.Run("returns paths for opencode when configs exist", func(t *testing.T) {
+		home, _ := os.UserHomeDir()
+		configDir := filepath.Join(home, ".config", "opencode")
+		stateDir := filepath.Join(home, ".local", "state", "opencode")
+		os.MkdirAll(configDir, 0o755)
+		os.MkdirAll(stateDir, 0o755)
+		os.WriteFile(filepath.Join(configDir, "opencode.json"), []byte(`{}`), 0o644)
+		os.WriteFile(filepath.Join(stateDir, "model.json"), []byte(`{}`), 0o644)
+
+		r := integrations["opencode"]
+		paths := editorPaths(r)
+		if len(paths) != 2 {
+			t.Errorf("expected 2 paths, got %d: %v", len(paths), paths)
+		}
+	})
+}
+
+func TestLoadIntegration_CorruptedJSON(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Create corrupted config.json file
+	dir := filepath.Join(tmpDir, ".ollama", "config")
+	os.MkdirAll(dir, 0o755)
+	os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{corrupted json`), 0o644)
+
+	// Corrupted file is treated as empty, so loadIntegration returns not found
+	_, err := loadIntegration("test")
+	if err == nil {
+		t.Error("expected error for nonexistent integration in corrupted file")
+	}
+}
+
+func TestSaveIntegration_NilModels(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	if err := saveIntegration("test", nil); err != nil {
+		t.Fatalf("saveIntegration with nil models failed: %v", err)
+	}
+
+	config, err := loadIntegration("test")
+	if err != nil {
+		t.Fatalf("loadIntegration failed: %v", err)
+	}
+
+	if config.Models == nil {
+		// nil is acceptable
+	} else if len(config.Models) != 0 {
+		t.Errorf("expected empty or nil models, got %v", config.Models)
+	}
+}
+
+func TestSaveIntegration_EmptyAppName(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	err := saveIntegration("", []string{"model"})
+	if err == nil {
+		t.Error("expected error for empty app name, got nil")
+	}
+	if err != nil && !strings.Contains(err.Error(), "app name cannot be empty") {
+		t.Errorf("expected 'app name cannot be empty' error, got: %v", err)
+	}
+}
+
+func TestLoadIntegration_NonexistentIntegration(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	_, err := loadIntegration("nonexistent")
+	if err == nil {
+		t.Error("expected error for nonexistent integration, got nil")
+	}
+	if !os.IsNotExist(err) {
+		t.Logf("error type is os.ErrNotExist as expected: %v", err)
+	}
+}
+
+func TestConfigPath(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	path, err := configPath()
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	expected := filepath.Join(tmpDir, ".ollama", "config", "config.json")
+	if path != expected {
+		t.Errorf("expected %s, got %s", expected, path)
+	}
+}
+
+func TestLoad(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("returns empty config when file does not exist", func(t *testing.T) {
+		cfg, err := load()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if cfg == nil {
+			t.Fatal("expected non-nil config")
+		}
+		if cfg.Integrations == nil {
+			t.Error("expected non-nil Integrations map")
+		}
+		if len(cfg.Integrations) != 0 {
+			t.Errorf("expected empty Integrations, got %d", len(cfg.Integrations))
+		}
+	})
+
+	t.Run("loads existing config", func(t *testing.T) {
+		path, _ := configPath()
+		os.MkdirAll(filepath.Dir(path), 0o755)
+		os.WriteFile(path, []byte(`{"integrations":{"test":{"models":["model-a"]}}}`), 0o644)
+
+		cfg, err := load()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if cfg.Integrations["test"] == nil {
+			t.Fatal("expected test integration")
+		}
+		if len(cfg.Integrations["test"].Models) != 1 {
+			t.Errorf("expected 1 model, got %d", len(cfg.Integrations["test"].Models))
+		}
+	})
+
+	t.Run("returns error for corrupted JSON", func(t *testing.T) {
+		path, _ := configPath()
+		os.MkdirAll(filepath.Dir(path), 0o755)
+		os.WriteFile(path, []byte(`{corrupted`), 0o644)
+
+		_, err := load()
+		if err == nil {
+			t.Error("expected error for corrupted JSON")
+		}
+	})
+}
+
+func TestSave(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("creates config file", func(t *testing.T) {
+		cfg := &config{
+			Integrations: map[string]*integration{
+				"test": {Models: []string{"model-a", "model-b"}},
+			},
+		}
+
+		if err := save(cfg); err != nil {
+			t.Fatal(err)
+		}
+
+		path, _ := configPath()
+		if _, err := os.Stat(path); os.IsNotExist(err) {
+			t.Error("config file was not created")
+		}
+	})
+
+	t.Run("round-trip preserves data", func(t *testing.T) {
+		cfg := &config{
+			Integrations: map[string]*integration{
+				"claude": {Models: []string{"llama3.2", "mistral"}},
+				"codex":  {Models: []string{"qwen2.5"}},
+			},
+		}
+
+		if err := save(cfg); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, err := load()
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if len(loaded.Integrations) != 2 {
+			t.Errorf("expected 2 integrations, got %d", len(loaded.Integrations))
+		}
+		if loaded.Integrations["claude"] == nil {
+			t.Error("missing claude integration")
+		}
+		if len(loaded.Integrations["claude"].Models) != 2 {
+			t.Errorf("expected 2 models for claude, got %d", len(loaded.Integrations["claude"].Models))
+		}
+	})
+}

cmd/config/droid.go

@@ -0,0 +1,164 @@
+package config
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"slices"
+)
+
+// Droid implements Runner and Editor for Droid integration
+type Droid struct{}
+
+// droidModelEntry represents a custom model entry in Droid's settings.json
+type droidModelEntry struct {
+	Model           string `json:"model"`
+	DisplayName     string `json:"displayName"`
+	BaseURL         string `json:"baseUrl"`
+	APIKey          string `json:"apiKey"`
+	Provider        string `json:"provider"`
+	MaxOutputTokens int    `json:"maxOutputTokens"`
+	SupportsImages  bool   `json:"supportsImages"`
+	ID              string `json:"id"`
+	Index           int    `json:"index"`
+}
+
+func (d *Droid) String() string { return "Droid" }
+
+func (d *Droid) Run(model string) error {
+	if _, err := exec.LookPath("droid"); err != nil {
+		return fmt.Errorf("droid is not installed, install from https://docs.factory.ai/cli/getting-started/quickstart")
+	}
+
+	// Call Edit() to ensure config is up-to-date before launch
+	models := []string{model}
+	if config, err := loadIntegration("droid"); err == nil && len(config.Models) > 0 {
+		models = config.Models
+	}
+	if err := d.Edit(models); err != nil {
+		return fmt.Errorf("setup failed: %w", err)
+	}
+
+	cmd := exec.Command("droid")
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	return cmd.Run()
+}
+
+func (d *Droid) Paths() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+	p := filepath.Join(home, ".factory", "settings.json")
+	if _, err := os.Stat(p); err == nil {
+		return []string{p}
+	}
+	return nil
+}
+
+func (d *Droid) Edit(models []string) error {
+	if len(models) == 0 {
+		return nil
+	}
+
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return err
+	}
+
+	settingsPath := filepath.Join(home, ".factory", "settings.json")
+	if err := os.MkdirAll(filepath.Dir(settingsPath), 0o755); err != nil {
+		return err
+	}
+
+	settings := make(map[string]any)
+	if data, err := os.ReadFile(settingsPath); err == nil {
+		if err := json.Unmarshal(data, &settings); err != nil {
+			return fmt.Errorf("failed to parse settings file: %w, at: %s", err, settingsPath)
+		}
+	}
+
+	customModels, _ := settings["customModels"].([]any)
+
+	// Keep only non-Ollama models (we'll rebuild Ollama models fresh)
+	nonOllamaModels := slices.DeleteFunc(slices.Clone(customModels), func(m any) bool {
+		entry, ok := m.(droidModelEntry)
+		if !ok {
+			return false
+		}
+		return entry.APIKey != "ollama"
+	})
+
+	// Build new Ollama model entries with sequential indices (0, 1, 2, ...)
+	var ollamaModels []any
+	var defaultModelID string
+	for i, model := range models {
+		modelID := fmt.Sprintf("custom:%s-%d", model, i)
+		ollamaModels = append(ollamaModels, droidModelEntry{
+			Model:           model,
+			DisplayName:     model,
+			BaseURL:         "http://localhost:11434/v1",
+			APIKey:          "ollama",
+			Provider:        "generic-chat-completion-api",
+			MaxOutputTokens: 64000,
+			SupportsImages:  false,
+			ID:              modelID,
+			Index:           i,
+		})
+		if i == 0 {
+			defaultModelID = modelID
+		}
+	}
+
+	settings["customModels"] = append(ollamaModels, nonOllamaModels...)
+
+	sessionSettings, ok := settings["sessionDefaultSettings"].(map[string]any)
+	if !ok {
+		sessionSettings = make(map[string]any)
+	}
+	sessionSettings["model"] = defaultModelID
+
+	if effort, ok := sessionSettings["reasoningEffort"].(string); !ok || !isValidReasoningEffort(effort) {
+		sessionSettings["reasoningEffort"] = "none"
+	}
+
+	settings["sessionDefaultSettings"] = sessionSettings
+
+	data, err := json.MarshalIndent(settings, "", "  ")
+	if err != nil {
+		return err
+	}
+	return writeWithBackup(settingsPath, data)
+}
+
+func (d *Droid) Models() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+	settings, err := readJSONFile(filepath.Join(home, ".factory", "settings.json"))
+	if err != nil {
+		return nil
+	}
+
+	customModels, _ := settings["customModels"].([]droidModelEntry)
+
+	var result []string
+	for _, m := range customModels {
+		if m.APIKey != "ollama" {
+			continue
+		}
+		result = append(result, m.Model)
+	}
+	return result
+}
+
+var validReasoningEfforts = []string{"high", "medium", "low", "none"}
+
+func isValidReasoningEffort(effort string) bool {
+	return slices.Contains(validReasoningEfforts, effort)
+}

cmd/config/droid_test.go

@@ -0,0 +1,454 @@
+package config
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestDroidIntegration(t *testing.T) {
+	d := &Droid{}
+
+	t.Run("String", func(t *testing.T) {
+		if got := d.String(); got != "Droid" {
+			t.Errorf("String() = %q, want %q", got, "Droid")
+		}
+	})
+
+	t.Run("implements Runner", func(t *testing.T) {
+		var _ Runner = d
+	})
+
+	t.Run("implements Editor", func(t *testing.T) {
+		var _ Editor = d
+	})
+}
+
+func TestDroidEdit(t *testing.T) {
+	d := &Droid{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	settingsDir := filepath.Join(tmpDir, ".factory")
+	settingsPath := filepath.Join(settingsDir, "settings.json")
+
+	cleanup := func() {
+		os.RemoveAll(settingsDir)
+	}
+
+	readSettings := func() map[string]any {
+		data, _ := os.ReadFile(settingsPath)
+		var settings map[string]any
+		json.Unmarshal(data, &settings)
+		return settings
+	}
+
+	getCustomModels := func(settings map[string]any) []map[string]any {
+		models, ok := settings["customModels"].([]any)
+		if !ok {
+			return nil
+		}
+		var result []map[string]any
+		for _, m := range models {
+			if entry, ok := m.(map[string]any); ok {
+				result = append(result, entry)
+			}
+		}
+		return result
+	}
+
+	t.Run("fresh install creates models with sequential indices", func(t *testing.T) {
+		cleanup()
+		if err := d.Edit([]string{"model-a", "model-b"}); err != nil {
+			t.Fatal(err)
+		}
+
+		settings := readSettings()
+		models := getCustomModels(settings)
+
+		if len(models) != 2 {
+			t.Fatalf("expected 2 models, got %d", len(models))
+		}
+
+		// Check first model
+		if models[0]["model"] != "model-a" {
+			t.Errorf("expected model-a, got %s", models[0]["model"])
+		}
+		if models[0]["id"] != "custom:model-a-[Ollama]-0" {
+			t.Errorf("expected custom:model-a-[Ollama]-0, got %s", models[0]["id"])
+		}
+		if models[0]["index"] != float64(0) {
+			t.Errorf("expected index 0, got %v", models[0]["index"])
+		}
+
+		// Check second model
+		if models[1]["model"] != "model-b" {
+			t.Errorf("expected model-b, got %s", models[1]["model"])
+		}
+		if models[1]["id"] != "custom:model-b-[Ollama]-1" {
+			t.Errorf("expected custom:model-b-[Ollama]-1, got %s", models[1]["id"])
+		}
+		if models[1]["index"] != float64(1) {
+			t.Errorf("expected index 1, got %v", models[1]["index"])
+		}
+	})
+
+	t.Run("sets sessionDefaultSettings.model to first model ID", func(t *testing.T) {
+		cleanup()
+		if err := d.Edit([]string{"model-a", "model-b"}); err != nil {
+			t.Fatal(err)
+		}
+
+		settings := readSettings()
+		session, ok := settings["sessionDefaultSettings"].(map[string]any)
+		if !ok {
+			t.Fatal("sessionDefaultSettings not found")
+		}
+		if session["model"] != "custom:model-a-[Ollama]-0" {
+			t.Errorf("expected custom:model-a-[Ollama]-0, got %s", session["model"])
+		}
+	})
+
+	t.Run("re-indexes when models removed", func(t *testing.T) {
+		cleanup()
+		// Add three models
+		d.Edit([]string{"model-a", "model-b", "model-c"})
+
+		// Remove middle model
+		d.Edit([]string{"model-a", "model-c"})
+
+		settings := readSettings()
+		models := getCustomModels(settings)
+
+		if len(models) != 2 {
+			t.Fatalf("expected 2 models, got %d", len(models))
+		}
+
+		// Check indices are sequential 0, 1
+		if models[0]["index"] != float64(0) {
+			t.Errorf("expected index 0, got %v", models[0]["index"])
+		}
+		if models[1]["index"] != float64(1) {
+			t.Errorf("expected index 1, got %v", models[1]["index"])
+		}
+
+		// Check IDs match new indices
+		if models[0]["id"] != "custom:model-a-[Ollama]-0" {
+			t.Errorf("expected custom:model-a-[Ollama]-0, got %s", models[0]["id"])
+		}
+		if models[1]["id"] != "custom:model-c-[Ollama]-1" {
+			t.Errorf("expected custom:model-c-[Ollama]-1, got %s", models[1]["id"])
+		}
+	})
+
+	t.Run("preserves non-Ollama custom models", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(settingsDir, 0o755)
+		// Pre-existing non-Ollama model
+		os.WriteFile(settingsPath, []byte(`{
+			"customModels": [
+				{"model": "gpt-4", "displayName": "GPT-4", "provider": "openai"}
+			]
+		}`), 0o644)
+
+		d.Edit([]string{"model-a"})
+
+		settings := readSettings()
+		models := getCustomModels(settings)
+
+		if len(models) != 2 {
+			t.Fatalf("expected 2 models (1 Ollama + 1 non-Ollama), got %d", len(models))
+		}
+
+		// Ollama model should be first
+		if models[0]["model"] != "model-a" {
+			t.Errorf("expected Ollama model first, got %s", models[0]["model"])
+		}
+
+		// Non-Ollama model should be preserved at end
+		if models[1]["model"] != "gpt-4" {
+			t.Errorf("expected gpt-4 preserved, got %s", models[1]["model"])
+		}
+	})
+
+	t.Run("preserves other settings", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(settingsDir, 0o755)
+		os.WriteFile(settingsPath, []byte(`{
+			"theme": "dark",
+			"enableHooks": true,
+			"sessionDefaultSettings": {"autonomyMode": "auto-high"}
+		}`), 0o644)
+
+		d.Edit([]string{"model-a"})
+
+		settings := readSettings()
+
+		if settings["theme"] != "dark" {
+			t.Error("theme was not preserved")
+		}
+		if settings["enableHooks"] != true {
+			t.Error("enableHooks was not preserved")
+		}
+
+		session := settings["sessionDefaultSettings"].(map[string]any)
+		if session["autonomyMode"] != "auto-high" {
+			t.Error("autonomyMode was not preserved")
+		}
+	})
+
+	t.Run("required fields present", func(t *testing.T) {
+		cleanup()
+		d.Edit([]string{"test-model"})
+
+		settings := readSettings()
+		models := getCustomModels(settings)
+
+		if len(models) != 1 {
+			t.Fatal("expected 1 model")
+		}
+
+		model := models[0]
+		requiredFields := []string{"model", "displayName", "baseUrl", "apiKey", "provider", "maxOutputTokens", "id", "index"}
+		for _, field := range requiredFields {
+			if model[field] == nil {
+				t.Errorf("missing required field: %s", field)
+			}
+		}
+
+		if model["baseUrl"] != "http://localhost:11434/v1" {
+			t.Errorf("unexpected baseUrl: %s", model["baseUrl"])
+		}
+		if model["apiKey"] != "ollama" {
+			t.Errorf("unexpected apiKey: %s", model["apiKey"])
+		}
+		if model["provider"] != "generic-chat-completion-api" {
+			t.Errorf("unexpected provider: %s", model["provider"])
+		}
+	})
+
+	t.Run("fixes invalid reasoningEffort", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(settingsDir, 0o755)
+		// Pre-existing settings with invalid reasoningEffort
+		os.WriteFile(settingsPath, []byte(`{
+			"sessionDefaultSettings": {"reasoningEffort": "off"}
+		}`), 0o644)
+
+		d.Edit([]string{"model-a"})
+
+		settings := readSettings()
+		session := settings["sessionDefaultSettings"].(map[string]any)
+
+		if session["reasoningEffort"] != "none" {
+			t.Errorf("expected reasoningEffort to be fixed to 'none', got %s", session["reasoningEffort"])
+		}
+	})
+
+	t.Run("preserves valid reasoningEffort", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(settingsDir, 0o755)
+		os.WriteFile(settingsPath, []byte(`{
+			"sessionDefaultSettings": {"reasoningEffort": "high"}
+		}`), 0o644)
+
+		d.Edit([]string{"model-a"})
+
+		settings := readSettings()
+		session := settings["sessionDefaultSettings"].(map[string]any)
+
+		if session["reasoningEffort"] != "high" {
+			t.Errorf("expected reasoningEffort to remain 'high', got %s", session["reasoningEffort"])
+		}
+	})
+}
+
+// Edge case tests for droid.go
+
+func TestDroidEdit_CorruptedJSON(t *testing.T) {
+	d := &Droid{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	settingsDir := filepath.Join(tmpDir, ".factory")
+	settingsPath := filepath.Join(settingsDir, "settings.json")
+
+	os.MkdirAll(settingsDir, 0o755)
+	os.WriteFile(settingsPath, []byte(`{corrupted json content`), 0o644)
+
+	// Corrupted JSON should return an error so user knows something is wrong
+	err := d.Edit([]string{"model-a"})
+	if err == nil {
+		t.Fatal("expected error for corrupted JSON, got nil")
+	}
+
+	// Original corrupted file should be preserved (not overwritten)
+	data, _ := os.ReadFile(settingsPath)
+	if string(data) != `{corrupted json content` {
+		t.Errorf("corrupted file was modified: got %s", string(data))
+	}
+}
+
+func TestDroidEdit_WrongTypeCustomModels(t *testing.T) {
+	d := &Droid{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	settingsDir := filepath.Join(tmpDir, ".factory")
+	settingsPath := filepath.Join(settingsDir, "settings.json")
+
+	os.MkdirAll(settingsDir, 0o755)
+	// customModels is a string instead of array
+	os.WriteFile(settingsPath, []byte(`{"customModels": "not an array"}`), 0o644)
+
+	// Should not panic - wrong type should be handled gracefully
+	err := d.Edit([]string{"model-a"})
+	if err != nil {
+		t.Fatalf("Edit failed with wrong type customModels: %v", err)
+	}
+
+	// Verify models were added correctly
+	data, _ := os.ReadFile(settingsPath)
+	var settings map[string]any
+	json.Unmarshal(data, &settings)
+
+	customModels, ok := settings["customModels"].([]any)
+	if !ok {
+		t.Fatalf("customModels should be array after setup, got %T", settings["customModels"])
+	}
+	if len(customModels) != 1 {
+		t.Errorf("expected 1 model, got %d", len(customModels))
+	}
+}
+
+func TestDroidEdit_EmptyModels(t *testing.T) {
+	d := &Droid{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	settingsDir := filepath.Join(tmpDir, ".factory")
+	settingsPath := filepath.Join(settingsDir, "settings.json")
+
+	os.MkdirAll(settingsDir, 0o755)
+	originalContent := `{"customModels": [{"model": "existing"}]}`
+	os.WriteFile(settingsPath, []byte(originalContent), 0o644)
+
+	// Empty models should be no-op
+	err := d.Edit([]string{})
+	if err != nil {
+		t.Fatalf("Edit with empty models failed: %v", err)
+	}
+
+	// Original content should be preserved (file not modified)
+	data, _ := os.ReadFile(settingsPath)
+	if string(data) != originalContent {
+		t.Errorf("empty models should not modify file, but content changed")
+	}
+}
+
+func TestDroidEdit_DuplicateModels(t *testing.T) {
+	d := &Droid{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	settingsDir := filepath.Join(tmpDir, ".factory")
+	settingsPath := filepath.Join(settingsDir, "settings.json")
+
+	// Add same model twice
+	err := d.Edit([]string{"model-a", "model-a"})
+	if err != nil {
+		t.Fatalf("Edit with duplicates failed: %v", err)
+	}
+
+	settings, err := readJSONFile(settingsPath)
+	if err != nil {
+		t.Fatalf("readJSONFile failed: %v", err)
+	}
+
+	customModels, _ := settings["customModels"].([]any)
+	// Document current behavior: duplicates are kept as separate entries
+	if len(customModels) != 2 {
+		t.Logf("Note: duplicates result in %d entries (documenting behavior)", len(customModels))
+	}
+}
+
+func TestDroidEdit_MalformedModelEntry(t *testing.T) {
+	d := &Droid{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	settingsDir := filepath.Join(tmpDir, ".factory")
+	settingsPath := filepath.Join(settingsDir, "settings.json")
+
+	os.MkdirAll(settingsDir, 0o755)
+	// Model entry is a string instead of a map
+	os.WriteFile(settingsPath, []byte(`{"customModels": ["not a map", 123]}`), 0o644)
+
+	err := d.Edit([]string{"model-a"})
+	if err != nil {
+		t.Fatalf("Edit with malformed entries failed: %v", err)
+	}
+
+	// Malformed entries should be preserved in nonOllamaModels
+	settings, _ := readJSONFile(settingsPath)
+	customModels, _ := settings["customModels"].([]any)
+
+	// Should have: 1 new Ollama model + 2 preserved malformed entries
+	if len(customModels) != 3 {
+		t.Errorf("expected 3 entries (1 new + 2 preserved malformed), got %d", len(customModels))
+	}
+}
+
+func TestDroidEdit_WrongTypeSessionSettings(t *testing.T) {
+	d := &Droid{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	settingsDir := filepath.Join(tmpDir, ".factory")
+	settingsPath := filepath.Join(settingsDir, "settings.json")
+
+	os.MkdirAll(settingsDir, 0o755)
+	// sessionDefaultSettings is a string instead of map
+	os.WriteFile(settingsPath, []byte(`{"sessionDefaultSettings": "not a map"}`), 0o644)
+
+	err := d.Edit([]string{"model-a"})
+	if err != nil {
+		t.Fatalf("Edit with wrong type sessionDefaultSettings failed: %v", err)
+	}
+
+	// Should create proper sessionDefaultSettings
+	settings, _ := readJSONFile(settingsPath)
+	session, ok := settings["sessionDefaultSettings"].(map[string]any)
+	if !ok {
+		t.Fatalf("sessionDefaultSettings should be map after setup, got %T", settings["sessionDefaultSettings"])
+	}
+	if session["model"] == nil {
+		t.Error("expected model to be set in sessionDefaultSettings")
+	}
+}
+
+func TestIsValidReasoningEffort(t *testing.T) {
+	tests := []struct {
+		effort string
+		valid  bool
+	}{
+		{"high", true},
+		{"medium", true},
+		{"low", true},
+		{"none", true},
+		{"off", false},
+		{"", false},
+		{"HIGH", false}, // case sensitive
+		{"max", false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.effort, func(t *testing.T) {
+			got := isValidReasoningEffort(tt.effort)
+			if got != tt.valid {
+				t.Errorf("isValidReasoningEffort(%q) = %v, want %v", tt.effort, got, tt.valid)
+			}
+		})
+	}
+}

cmd/config/files.go

@@ -0,0 +1,99 @@
+package config
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"time"
+)
+
+func readJSONFile(path string) (map[string]any, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, err
+	}
+	var result map[string]any
+	if err := json.Unmarshal(data, &result); err != nil {
+		return nil, err
+	}
+	return result, nil
+}
+
+func copyFile(src, dst string) error {
+	info, err := os.Stat(src)
+	if err != nil {
+		return err
+	}
+	data, err := os.ReadFile(src)
+	if err != nil {
+		return err
+	}
+	return os.WriteFile(dst, data, info.Mode().Perm())
+}
+
+func backupDir() string {
+	return filepath.Join(os.TempDir(), "ollama-backups")
+}
+
+func backupToTmp(srcPath string) (string, error) {
+	dir := backupDir()
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return "", err
+	}
+
+	backupPath := filepath.Join(dir, fmt.Sprintf("%s.%d", filepath.Base(srcPath), time.Now().Unix()))
+	if err := copyFile(srcPath, backupPath); err != nil {
+		return "", err
+	}
+	return backupPath, nil
+}
+
+// writeWithBackup writes data to path via temp file + rename, backing up any existing file first
+func writeWithBackup(path string, data []byte) error {
+	var backupPath string
+	// backup must be created before any writes to the target file
+	if existingContent, err := os.ReadFile(path); err == nil {
+		if !bytes.Equal(existingContent, data) {
+			backupPath, err = backupToTmp(path)
+			if err != nil {
+				return fmt.Errorf("backup failed: %w", err)
+			}
+		}
+	} else if !os.IsNotExist(err) {
+		return fmt.Errorf("read existing file: %w", err)
+	}
+
+	dir := filepath.Dir(path)
+	tmp, err := os.CreateTemp(dir, ".tmp-*")
+	if err != nil {
+		return fmt.Errorf("create temp failed: %w", err)
+	}
+	tmpPath := tmp.Name()
+
+	if _, err := tmp.Write(data); err != nil {
+		_ = tmp.Close()
+		_ = os.Remove(tmpPath)
+		return fmt.Errorf("write failed: %w", err)
+	}
+	if err := tmp.Sync(); err != nil {
+		_ = tmp.Close()
+		_ = os.Remove(tmpPath)
+		return fmt.Errorf("sync failed: %w", err)
+	}
+	if err := tmp.Close(); err != nil {
+		_ = os.Remove(tmpPath)
+		return fmt.Errorf("close failed: %w", err)
+	}
+
+	if err := os.Rename(tmpPath, path); err != nil {
+		_ = os.Remove(tmpPath)
+		if backupPath != "" {
+			_ = copyFile(backupPath, path)
+		}
+		return fmt.Errorf("rename failed: %w", err)
+	}
+
+	return nil
+}

cmd/config/files_test.go

@@ -0,0 +1,502 @@
+package config
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"runtime"
+	"testing"
+)
+
+func mustMarshal(t *testing.T, v any) []byte {
+	t.Helper()
+	data, err := json.MarshalIndent(v, "", "  ")
+	if err != nil {
+		t.Fatal(err)
+	}
+	return data
+}
+
+func TestWriteWithBackup(t *testing.T) {
+	tmpDir := t.TempDir()
+
+	t.Run("creates file", func(t *testing.T) {
+		path := filepath.Join(tmpDir, "new.json")
+		data := mustMarshal(t, map[string]string{"key": "value"})
+
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		content, err := os.ReadFile(path)
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		var result map[string]string
+		if err := json.Unmarshal(content, &result); err != nil {
+			t.Fatal(err)
+		}
+		if result["key"] != "value" {
+			t.Errorf("expected value, got %s", result["key"])
+		}
+	})
+
+	t.Run("creates backup in /tmp/ollama-backups", func(t *testing.T) {
+		path := filepath.Join(tmpDir, "backup.json")
+
+		os.WriteFile(path, []byte(`{"original": true}`), 0o644)
+
+		data := mustMarshal(t, map[string]bool{"updated": true})
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		entries, err := os.ReadDir(backupDir())
+		if err != nil {
+			t.Fatal("backup directory not created")
+		}
+
+		var foundBackup bool
+		for _, entry := range entries {
+			if filepath.Ext(entry.Name()) != ".json" {
+				name := entry.Name()
+				if len(name) > len("backup.json.") && name[:len("backup.json.")] == "backup.json." {
+					backupPath := filepath.Join(backupDir(), name)
+					backup, err := os.ReadFile(backupPath)
+					if err == nil {
+						var backupData map[string]bool
+						json.Unmarshal(backup, &backupData)
+						if backupData["original"] {
+							foundBackup = true
+							os.Remove(backupPath)
+							break
+						}
+					}
+				}
+			}
+		}
+
+		if !foundBackup {
+			t.Error("backup file not created in /tmp/ollama-backups")
+		}
+
+		current, _ := os.ReadFile(path)
+		var currentData map[string]bool
+		json.Unmarshal(current, &currentData)
+		if !currentData["updated"] {
+			t.Error("file doesn't contain updated data")
+		}
+	})
+
+	t.Run("no backup for new file", func(t *testing.T) {
+		path := filepath.Join(tmpDir, "nobak.json")
+
+		data := mustMarshal(t, map[string]string{"new": "file"})
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		entries, _ := os.ReadDir(backupDir())
+		for _, entry := range entries {
+			if len(entry.Name()) > len("nobak.json.") && entry.Name()[:len("nobak.json.")] == "nobak.json." {
+				t.Error("backup should not exist for new file")
+			}
+		}
+	})
+
+	t.Run("no backup when content unchanged", func(t *testing.T) {
+		path := filepath.Join(tmpDir, "unchanged.json")
+
+		data := mustMarshal(t, map[string]string{"key": "value"})
+
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		entries1, _ := os.ReadDir(backupDir())
+		countBefore := 0
+		for _, e := range entries1 {
+			if len(e.Name()) > len("unchanged.json.") && e.Name()[:len("unchanged.json.")] == "unchanged.json." {
+				countBefore++
+			}
+		}
+
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		entries2, _ := os.ReadDir(backupDir())
+		countAfter := 0
+		for _, e := range entries2 {
+			if len(e.Name()) > len("unchanged.json.") && e.Name()[:len("unchanged.json.")] == "unchanged.json." {
+				countAfter++
+			}
+		}
+
+		if countAfter != countBefore {
+			t.Errorf("backup was created when content unchanged (before=%d, after=%d)", countBefore, countAfter)
+		}
+	})
+
+	t.Run("backup filename contains unix timestamp", func(t *testing.T) {
+		path := filepath.Join(tmpDir, "timestamped.json")
+
+		os.WriteFile(path, []byte(`{"v": 1}`), 0o644)
+		data := mustMarshal(t, map[string]int{"v": 2})
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatal(err)
+		}
+
+		entries, _ := os.ReadDir(backupDir())
+		var found bool
+		for _, entry := range entries {
+			name := entry.Name()
+			if len(name) > len("timestamped.json.") && name[:len("timestamped.json.")] == "timestamped.json." {
+				timestamp := name[len("timestamped.json."):]
+				for _, c := range timestamp {
+					if c < '0' || c > '9' {
+						t.Errorf("backup filename timestamp contains non-numeric character: %s", name)
+					}
+				}
+				found = true
+				os.Remove(filepath.Join(backupDir(), name))
+				break
+			}
+		}
+		if !found {
+			t.Error("backup file with timestamp not found")
+		}
+	})
+}
+
+// Edge case tests for files.go
+
+// TestWriteWithBackup_FailsIfBackupFails documents critical behavior: if backup fails, we must not proceed.
+// User could lose their config with no way to recover.
+func TestWriteWithBackup_FailsIfBackupFails(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("permission tests unreliable on Windows")
+	}
+
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "config.json")
+
+	// Create original file
+	originalContent := []byte(`{"original": true}`)
+	os.WriteFile(path, originalContent, 0o644)
+
+	// Make backup directory read-only to force backup failure
+	backupDir := backupDir()
+	os.MkdirAll(backupDir, 0o755)
+	os.Chmod(backupDir, 0o444) // Read-only
+	defer os.Chmod(backupDir, 0o755)
+
+	newContent := []byte(`{"updated": true}`)
+	err := writeWithBackup(path, newContent)
+
+	// Should fail because backup couldn't be created
+	if err == nil {
+		t.Error("expected error when backup fails, got nil")
+	}
+
+	// Original file should be preserved
+	current, _ := os.ReadFile(path)
+	if string(current) != string(originalContent) {
+		t.Errorf("original file was modified despite backup failure: got %s", string(current))
+	}
+}
+
+// TestWriteWithBackup_PermissionDenied verifies clear error when target file has wrong permissions.
+// Common issue when config owned by root or wrong perms.
+func TestWriteWithBackup_PermissionDenied(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("permission tests unreliable on Windows")
+	}
+
+	tmpDir := t.TempDir()
+
+	// Create a read-only directory
+	readOnlyDir := filepath.Join(tmpDir, "readonly")
+	os.MkdirAll(readOnlyDir, 0o755)
+	os.Chmod(readOnlyDir, 0o444)
+	defer os.Chmod(readOnlyDir, 0o755)
+
+	path := filepath.Join(readOnlyDir, "config.json")
+	err := writeWithBackup(path, []byte(`{"test": true}`))
+
+	if err == nil {
+		t.Error("expected permission error, got nil")
+	}
+}
+
+// TestWriteWithBackup_DirectoryDoesNotExist verifies behavior when target directory doesn't exist.
+// writeWithBackup doesn't create directories - caller is responsible.
+func TestWriteWithBackup_DirectoryDoesNotExist(t *testing.T) {
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "nonexistent", "subdir", "config.json")
+
+	err := writeWithBackup(path, []byte(`{"test": true}`))
+
+	// Should fail because directory doesn't exist
+	if err == nil {
+		t.Error("expected error for nonexistent directory, got nil")
+	}
+}
+
+// TestWriteWithBackup_SymlinkTarget documents behavior when target is a symlink.
+// Documents what happens if user symlinks their config file.
+func TestWriteWithBackup_SymlinkTarget(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("symlink tests may require admin on Windows")
+	}
+
+	tmpDir := t.TempDir()
+	realFile := filepath.Join(tmpDir, "real.json")
+	symlink := filepath.Join(tmpDir, "link.json")
+
+	// Create real file and symlink
+	os.WriteFile(realFile, []byte(`{"v": 1}`), 0o644)
+	os.Symlink(realFile, symlink)
+
+	// Write through symlink
+	err := writeWithBackup(symlink, []byte(`{"v": 2}`))
+	if err != nil {
+		t.Fatalf("writeWithBackup through symlink failed: %v", err)
+	}
+
+	// The real file should be updated (symlink followed for temp file creation)
+	content, _ := os.ReadFile(symlink)
+	if string(content) != `{"v": 2}` {
+		t.Errorf("symlink target not updated correctly: got %s", string(content))
+	}
+}
+
+// TestBackupToTmp_SpecialCharsInFilename verifies backup works with special characters.
+// User may have config files with unusual names.
+func TestBackupToTmp_SpecialCharsInFilename(t *testing.T) {
+	tmpDir := t.TempDir()
+
+	// File with spaces and special chars
+	path := filepath.Join(tmpDir, "my config (backup).json")
+	os.WriteFile(path, []byte(`{"test": true}`), 0o644)
+
+	backupPath, err := backupToTmp(path)
+	if err != nil {
+		t.Fatalf("backupToTmp with special chars failed: %v", err)
+	}
+
+	// Verify backup exists and has correct content
+	content, err := os.ReadFile(backupPath)
+	if err != nil {
+		t.Fatalf("could not read backup: %v", err)
+	}
+	if string(content) != `{"test": true}` {
+		t.Errorf("backup content mismatch: got %s", string(content))
+	}
+
+	os.Remove(backupPath)
+}
+
+// TestCopyFile_PreservesPermissions verifies that copyFile preserves file permissions.
+func TestCopyFile_PreservesPermissions(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("permission preservation tests unreliable on Windows")
+	}
+
+	tmpDir := t.TempDir()
+	src := filepath.Join(tmpDir, "src.json")
+	dst := filepath.Join(tmpDir, "dst.json")
+
+	// Create source with specific permissions
+	os.WriteFile(src, []byte(`{"test": true}`), 0o600)
+
+	err := copyFile(src, dst)
+	if err != nil {
+		t.Fatalf("copyFile failed: %v", err)
+	}
+
+	srcInfo, _ := os.Stat(src)
+	dstInfo, _ := os.Stat(dst)
+
+	if srcInfo.Mode().Perm() != dstInfo.Mode().Perm() {
+		t.Errorf("permissions not preserved: src=%v, dst=%v", srcInfo.Mode().Perm(), dstInfo.Mode().Perm())
+	}
+}
+
+// TestCopyFile_SourceNotFound verifies clear error when source doesn't exist.
+func TestCopyFile_SourceNotFound(t *testing.T) {
+	tmpDir := t.TempDir()
+	src := filepath.Join(tmpDir, "nonexistent.json")
+	dst := filepath.Join(tmpDir, "dst.json")
+
+	err := copyFile(src, dst)
+	if err == nil {
+		t.Error("expected error for nonexistent source, got nil")
+	}
+}
+
+// TestWriteWithBackup_TargetIsDirectory verifies error when path points to a directory.
+func TestWriteWithBackup_TargetIsDirectory(t *testing.T) {
+	tmpDir := t.TempDir()
+	dirPath := filepath.Join(tmpDir, "actualdir")
+	os.MkdirAll(dirPath, 0o755)
+
+	err := writeWithBackup(dirPath, []byte(`{"test": true}`))
+	if err == nil {
+		t.Error("expected error when target is a directory, got nil")
+	}
+}
+
+// TestWriteWithBackup_EmptyData verifies writing zero bytes works correctly.
+func TestWriteWithBackup_EmptyData(t *testing.T) {
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "empty.json")
+
+	err := writeWithBackup(path, []byte{})
+	if err != nil {
+		t.Fatalf("writeWithBackup with empty data failed: %v", err)
+	}
+
+	content, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("could not read file: %v", err)
+	}
+	if len(content) != 0 {
+		t.Errorf("expected empty file, got %d bytes", len(content))
+	}
+}
+
+// TestWriteWithBackup_FileUnreadableButDirWritable verifies behavior when existing file
+// cannot be read (for backup comparison) but directory is writable.
+func TestWriteWithBackup_FileUnreadableButDirWritable(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("permission tests unreliable on Windows")
+	}
+
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "unreadable.json")
+
+	// Create file and make it unreadable
+	os.WriteFile(path, []byte(`{"original": true}`), 0o644)
+	os.Chmod(path, 0o000)
+	defer os.Chmod(path, 0o644)
+
+	// Should fail because we can't read the file to compare/backup
+	err := writeWithBackup(path, []byte(`{"updated": true}`))
+	if err == nil {
+		t.Error("expected error when file is unreadable, got nil")
+	}
+}
+
+// TestWriteWithBackup_RapidSuccessiveWrites verifies backup works with multiple writes
+// within the same second (timestamp collision scenario).
+func TestWriteWithBackup_RapidSuccessiveWrites(t *testing.T) {
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "rapid.json")
+
+	// Create initial file
+	os.WriteFile(path, []byte(`{"v": 0}`), 0o644)
+
+	// Rapid successive writes
+	for i := 1; i <= 3; i++ {
+		data := []byte(fmt.Sprintf(`{"v": %d}`, i))
+		if err := writeWithBackup(path, data); err != nil {
+			t.Fatalf("write %d failed: %v", i, err)
+		}
+	}
+
+	// Verify final content
+	content, _ := os.ReadFile(path)
+	if string(content) != `{"v": 3}` {
+		t.Errorf("expected final content {\"v\": 3}, got %s", string(content))
+	}
+
+	// Verify at least one backup exists
+	entries, _ := os.ReadDir(backupDir())
+	var backupCount int
+	for _, e := range entries {
+		if len(e.Name()) > len("rapid.json.") && e.Name()[:len("rapid.json.")] == "rapid.json." {
+			backupCount++
+		}
+	}
+	if backupCount == 0 {
+		t.Error("expected at least one backup file from rapid writes")
+	}
+}
+
+// TestWriteWithBackup_BackupDirIsFile verifies error when backup directory path is a file.
+func TestWriteWithBackup_BackupDirIsFile(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("test modifies system temp directory")
+	}
+
+	// Create a file at the backup directory path
+	backupPath := backupDir()
+	// Clean up any existing directory first
+	os.RemoveAll(backupPath)
+	// Create a file instead of directory
+	os.WriteFile(backupPath, []byte("not a directory"), 0o644)
+	defer func() {
+		os.Remove(backupPath)
+		os.MkdirAll(backupPath, 0o755)
+	}()
+
+	tmpDir := t.TempDir()
+	path := filepath.Join(tmpDir, "test.json")
+	os.WriteFile(path, []byte(`{"original": true}`), 0o644)
+
+	err := writeWithBackup(path, []byte(`{"updated": true}`))
+	if err == nil {
+		t.Error("expected error when backup dir is a file, got nil")
+	}
+}
+
+// TestWriteWithBackup_NoOrphanTempFiles verifies temp files are cleaned up on failure.
+func TestWriteWithBackup_NoOrphanTempFiles(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("permission tests unreliable on Windows")
+	}
+
+	tmpDir := t.TempDir()
+
+	// Count existing temp files
+	countTempFiles := func() int {
+		entries, _ := os.ReadDir(tmpDir)
+		count := 0
+		for _, e := range entries {
+			if len(e.Name()) > 4 && e.Name()[:4] == ".tmp" {
+				count++
+			}
+		}
+		return count
+	}
+
+	before := countTempFiles()
+
+	// Create a file, then make directory read-only to cause rename failure
+	path := filepath.Join(tmpDir, "orphan.json")
+	os.WriteFile(path, []byte(`{"v": 1}`), 0o644)
+
+	// Make a subdirectory and try to write there after making parent read-only
+	subDir := filepath.Join(tmpDir, "subdir")
+	os.MkdirAll(subDir, 0o755)
+	subPath := filepath.Join(subDir, "config.json")
+	os.WriteFile(subPath, []byte(`{"v": 1}`), 0o644)
+
+	// Make subdir read-only after creating temp file would succeed but rename would fail
+	// This is tricky to test - the temp file is created in the same dir, so if we can't
+	// rename, we also couldn't create. Let's just verify normal failure cleanup works.
+
+	// Force a failure by making the target a directory
+	badPath := filepath.Join(tmpDir, "isdir")
+	os.MkdirAll(badPath, 0o755)
+
+	_ = writeWithBackup(badPath, []byte(`{"test": true}`))
+
+	after := countTempFiles()
+	if after > before {
+		t.Errorf("orphan temp files left behind: before=%d, after=%d", before, after)
+	}
+}

cmd/config/integrations.go

@@ -0,0 +1,361 @@
+package config
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"maps"
+	"os"
+	"os/exec"
+	"runtime"
+	"slices"
+	"strings"
+	"time"
+
+	"github.com/ollama/ollama/api"
+	"github.com/spf13/cobra"
+)
+
+// Runners execute the launching of a model with the integration - claude, codex
+// Editors can edit config files (supports multi-model selection) - opencode, droid
+// They are composable interfaces where in some cases an editor is also a runner - opencode, droid
+// Runner can run an integration with a model.
+
+type Runner interface {
+	Run(model string) error
+	// String returns the human-readable name of the integration
+	String() string
+}
+
+// Editor can edit config files (supports multi-model selection)
+type Editor interface {
+	// Paths returns the paths to the config files for the integration
+	Paths() []string
+	// Edit updates the config files for the integration with the given models
+	Edit(models []string) error
+	// Models returns the models currently configured for the integration
+	Models() []string
+}
+
+// integrations is the registry of available integrations.
+var integrations = map[string]Runner{
+	"claude":   &Claude{},
+	"codex":    &Codex{},
+	"droid":    &Droid{},
+	"opencode": &OpenCode{},
+}
+
+func selectIntegration() (string, error) {
+	if len(integrations) == 0 {
+		return "", fmt.Errorf("no integrations available")
+	}
+
+	names := slices.Sorted(maps.Keys(integrations))
+	var items []selectItem
+	for _, name := range names {
+		r := integrations[name]
+		description := r.String()
+		if conn, err := loadIntegration(name); err == nil && len(conn.Models) > 0 {
+			description = fmt.Sprintf("%s (%s)", r.String(), conn.Models[0])
+		}
+		items = append(items, selectItem{Name: name, Description: description})
+	}
+
+	return selectPrompt("Select integration:", items)
+}
+
+// selectModels lets the user select models for an integration
+func selectModels(ctx context.Context, name, current string) ([]string, error) {
+	r, ok := integrations[name]
+	if !ok {
+		return nil, fmt.Errorf("unknown integration: %s", name)
+	}
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return nil, err
+	}
+
+	models, err := client.List(ctx)
+	if err != nil {
+		return nil, err
+	}
+
+	if len(models.Models) == 0 {
+		return nil, fmt.Errorf("no models available, run 'ollama pull <model>' first")
+	}
+
+	var items []selectItem
+	cloudModels := make(map[string]bool)
+	for _, m := range models.Models {
+		if m.RemoteModel != "" {
+			cloudModels[m.Name] = true
+		}
+		items = append(items, selectItem{Name: m.Name})
+	}
+
+	if len(items) == 0 {
+		return nil, fmt.Errorf("no local models available, run 'ollama pull <model>' first")
+	}
+
+	// Get previously configured models (saved config takes precedence)
+	var preChecked []string
+	if saved, err := loadIntegration(name); err == nil {
+		preChecked = saved.Models
+	} else if editor, ok := r.(Editor); ok {
+		preChecked = editor.Models()
+	}
+	checked := make(map[string]bool, len(preChecked))
+	for _, n := range preChecked {
+		checked[n] = true
+	}
+
+	// Resolve current to full name (e.g., "llama3.2" -> "llama3.2:latest")
+	for _, item := range items {
+		if item.Name == current || strings.HasPrefix(item.Name, current+":") {
+			current = item.Name
+			break
+		}
+	}
+
+	// If current model is configured, move to front of preChecked
+	if checked[current] {
+		preChecked = append([]string{current}, slices.DeleteFunc(preChecked, func(m string) bool { return m == current })...)
+	}
+
+	// Sort: checked first, then alphabetical
+	slices.SortFunc(items, func(a, b selectItem) int {
+		ac, bc := checked[a.Name], checked[b.Name]
+		if ac != bc {
+			if ac {
+				return -1
+			}
+			return 1
+		}
+		return strings.Compare(strings.ToLower(a.Name), strings.ToLower(b.Name))
+	})
+
+	var selected []string
+	// only editors support multi-model selection
+	if _, ok := r.(Editor); ok {
+		selected, err = multiSelectPrompt(fmt.Sprintf("Select models for %s:", r), items, preChecked)
+		if err != nil {
+			return nil, err
+		}
+	} else {
+		model, err := selectPrompt(fmt.Sprintf("Select model for %s:", r), items)
+		if err != nil {
+			return nil, err
+		}
+		selected = []string{model}
+	}
+
+	// if any model in selected is a cloud model, ensure signed in
+	var selectedCloudModels []string
+	for _, m := range selected {
+		if cloudModels[m] {
+			selectedCloudModels = append(selectedCloudModels, m)
+		}
+	}
+	if len(selectedCloudModels) > 0 {
+		// ensure user is signed in
+		user, err := client.Whoami(ctx)
+		if err == nil && user != nil && user.Name != "" {
+			return selected, nil
+		}
+
+		var aErr api.AuthorizationError
+		if !errors.As(err, &aErr) || aErr.SigninURL == "" {
+			return nil, err
+		}
+
+		modelList := strings.Join(selectedCloudModels, ", ")
+		yes, err := confirmPrompt(fmt.Sprintf("sign in to use %s?", modelList))
+		if err != nil || !yes {
+			return nil, fmt.Errorf("%s requires sign in", modelList)
+		}
+
+		fmt.Fprintf(os.Stderr, "\nTo sign in, navigate to:\n    %s\n\n", aErr.SigninURL)
+
+		// TODO(parthsareen): extract into auth package for cmd
+		// Auto-open browser (best effort, fail silently)
+		switch runtime.GOOS {
+		case "darwin":
+			_ = exec.Command("open", aErr.SigninURL).Start()
+		case "linux":
+			_ = exec.Command("xdg-open", aErr.SigninURL).Start()
+		case "windows":
+			_ = exec.Command("rundll32", "url.dll,FileProtocolHandler", aErr.SigninURL).Start()
+		}
+
+		spinnerFrames := []string{"|", "/", "-", "\\"}
+		frame := 0
+
+		fmt.Fprintf(os.Stderr, "\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[0])
+
+		ticker := time.NewTicker(200 * time.Millisecond)
+		defer ticker.Stop()
+
+		for {
+			select {
+			case <-ctx.Done():
+				fmt.Fprintf(os.Stderr, "\r\033[K")
+				return nil, ctx.Err()
+			case <-ticker.C:
+				frame++
+				fmt.Fprintf(os.Stderr, "\r\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[frame%len(spinnerFrames)])
+
+				// poll every 10th frame (~2 seconds)
+				if frame%10 == 0 {
+					u, err := client.Whoami(ctx)
+					if err == nil && u != nil && u.Name != "" {
+						fmt.Fprintf(os.Stderr, "\r\033[K\033[A\r\033[K\033[1msigned in:\033[0m %s\n", u.Name)
+						return selected, nil
+					}
+				}
+			}
+		}
+	}
+
+	return selected, nil
+}
+
+func runIntegration(name, modelName string) error {
+	r, ok := integrations[name]
+	if !ok {
+		return fmt.Errorf("unknown integration: %s", name)
+	}
+	fmt.Fprintf(os.Stderr, "\nLaunching %s with %s...\n", r, modelName)
+	return r.Run(modelName)
+}
+
+// ConfigCmd returns the cobra command for configuring integrations.
+func ConfigCmd(checkServerHeartbeat func(cmd *cobra.Command, args []string) error) *cobra.Command {
+	var modelFlag string
+	var launchFlag bool
+
+	cmd := &cobra.Command{
+		Use:   "config [INTEGRATION]",
+		Short: "Configure an external integration to use Ollama",
+		Long: `Configure an external application to use Ollama models.
+
+Supported integrations:
+  claude    Claude Code
+  codex     Codex
+  droid     Droid
+  opencode  OpenCode
+
+Examples:
+  ollama config
+  ollama config claude
+  ollama config droid --launch`,
+		Args:    cobra.MaximumNArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE: func(cmd *cobra.Command, args []string) error {
+			var name string
+			if len(args) > 0 {
+				name = args[0]
+			} else {
+				var err error
+				name, err = selectIntegration()
+				if errors.Is(err, errCancelled) {
+					return nil
+				}
+				if err != nil {
+					return err
+				}
+			}
+
+			r, ok := integrations[strings.ToLower(name)]
+			if !ok {
+				return fmt.Errorf("unknown integration: %s", name)
+			}
+
+			// If --launch without --model, use saved config if available
+			if launchFlag && modelFlag == "" {
+				if config, err := loadIntegration(name); err == nil && len(config.Models) > 0 {
+					return runIntegration(name, config.Models[0])
+				}
+			}
+
+			var models []string
+			if modelFlag != "" {
+				// When --model is specified, merge with existing models (new model becomes default)
+				models = []string{modelFlag}
+				if existing, err := loadIntegration(name); err == nil && len(existing.Models) > 0 {
+					for _, m := range existing.Models {
+						if m != modelFlag {
+							models = append(models, m)
+						}
+					}
+				}
+			} else {
+				var err error
+				models, err = selectModels(cmd.Context(), name, "")
+				if errors.Is(err, errCancelled) {
+					return nil
+				}
+				if err != nil {
+					return err
+				}
+			}
+
+			if editor, isEditor := r.(Editor); isEditor {
+				paths := editor.Paths()
+				if len(paths) > 0 {
+					fmt.Fprintf(os.Stderr, "This will modify your %s configuration:\n", r)
+					for _, p := range paths {
+						fmt.Fprintf(os.Stderr, "  %s\n", p)
+					}
+					fmt.Fprintf(os.Stderr, "Backups will be saved to %s/\n\n", backupDir())
+
+					if ok, _ := confirmPrompt("Proceed?"); !ok {
+						return nil
+					}
+				}
+			}
+
+			if err := saveIntegration(name, models); err != nil {
+				return fmt.Errorf("failed to save: %w", err)
+			}
+
+			if editor, isEditor := r.(Editor); isEditor {
+				if err := editor.Edit(models); err != nil {
+					return fmt.Errorf("setup failed: %w", err)
+				}
+			}
+
+			if _, isEditor := r.(Editor); isEditor {
+				if len(models) == 1 {
+					fmt.Fprintf(os.Stderr, "Added %s to %s\n", models[0], r)
+				} else {
+					fmt.Fprintf(os.Stderr, "Added %d models to %s (default: %s)\n", len(models), r, models[0])
+				}
+			}
+
+			if slices.ContainsFunc(models, func(m string) bool {
+				return !strings.HasSuffix(m, "cloud")
+			}) {
+				fmt.Fprintln(os.Stderr)
+				fmt.Fprintln(os.Stderr, "Coding agents work best with at least 64k context. Either:")
+				fmt.Fprintln(os.Stderr, "  - Set the context slider in Ollama app settings")
+				fmt.Fprintln(os.Stderr, "  - Run: OLLAMA_CONTEXT_LENGTH=64000 ollama serve")
+			}
+
+			if launchFlag {
+				return runIntegration(name, models[0])
+			}
+
+			if launch, _ := confirmPrompt(fmt.Sprintf("\nLaunch %s now?", r)); launch {
+				return runIntegration(name, models[0])
+			}
+
+			fmt.Fprintf(os.Stderr, "Run 'ollama config %s --launch' to start with %s\n", strings.ToLower(name), models[0])
+			return nil
+		},
+	}
+
+	cmd.Flags().StringVar(&modelFlag, "model", "", "Model to use")
+	cmd.Flags().BoolVar(&launchFlag, "launch", false, "Launch the integration after configuring")
+	return cmd
+}

cmd/config/integrations_test.go

@@ -0,0 +1,188 @@
+package config
+
+import (
+	"slices"
+	"strings"
+	"testing"
+
+	"github.com/spf13/cobra"
+)
+
+func TestIntegrationLookup(t *testing.T) {
+	tests := []struct {
+		name      string
+		input     string
+		wantFound bool
+		wantName  string
+	}{
+		{"claude lowercase", "claude", true, "Claude Code"},
+		{"claude uppercase", "CLAUDE", true, "Claude Code"},
+		{"claude mixed case", "Claude", true, "Claude Code"},
+		{"codex", "codex", true, "Codex"},
+		{"droid", "droid", true, "Droid"},
+		{"opencode", "opencode", true, "OpenCode"},
+		{"unknown integration", "unknown", false, ""},
+		{"empty string", "", false, ""},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			r, found := integrations[strings.ToLower(tt.input)]
+			if found != tt.wantFound {
+				t.Errorf("integrations[%q] found = %v, want %v", tt.input, found, tt.wantFound)
+			}
+			if found && r.String() != tt.wantName {
+				t.Errorf("integrations[%q].String() = %q, want %q", tt.input, r.String(), tt.wantName)
+			}
+		})
+	}
+}
+
+func TestIntegrationRegistry(t *testing.T) {
+	expectedIntegrations := []string{"claude", "codex", "droid", "opencode"}
+
+	for _, name := range expectedIntegrations {
+		t.Run(name, func(t *testing.T) {
+			r, ok := integrations[name]
+			if !ok {
+				t.Fatalf("integration %q not found in registry", name)
+			}
+			if r.String() == "" {
+				t.Error("integration.String() should not be empty")
+			}
+		})
+	}
+}
+
+func TestHasLocalModel(t *testing.T) {
+	tests := []struct {
+		name   string
+		models []string
+		want   bool
+	}{
+		{"empty list", []string{}, false},
+		{"single local model", []string{"llama3.2"}, true},
+		{"single cloud model", []string{"cloud-model"}, false},
+		{"mixed models", []string{"cloud-model", "llama3.2"}, true},
+		{"multiple local models", []string{"llama3.2", "qwen2.5"}, true},
+		{"multiple cloud models", []string{"cloud-a", "cloud-b"}, false},
+		{"local model first", []string{"llama3.2", "cloud-model"}, true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := slices.ContainsFunc(tt.models, func(m string) bool {
+				return !strings.Contains(m, "cloud")
+			})
+			if got != tt.want {
+				t.Errorf("hasLocalModel(%v) = %v, want %v", tt.models, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestConfigCmd(t *testing.T) {
+	// Mock checkServerHeartbeat that always succeeds
+	mockCheck := func(cmd *cobra.Command, args []string) error {
+		return nil
+	}
+
+	cmd := ConfigCmd(mockCheck)
+
+	t.Run("command structure", func(t *testing.T) {
+		if cmd.Use != "config [INTEGRATION]" {
+			t.Errorf("Use = %q, want %q", cmd.Use, "config [INTEGRATION]")
+		}
+		if cmd.Short == "" {
+			t.Error("Short description should not be empty")
+		}
+		if cmd.Long == "" {
+			t.Error("Long description should not be empty")
+		}
+	})
+
+	t.Run("flags exist", func(t *testing.T) {
+		modelFlag := cmd.Flags().Lookup("model")
+		if modelFlag == nil {
+			t.Error("--model flag should exist")
+		}
+
+		launchFlag := cmd.Flags().Lookup("launch")
+		if launchFlag == nil {
+			t.Error("--launch flag should exist")
+		}
+	})
+
+	t.Run("PreRunE is set", func(t *testing.T) {
+		if cmd.PreRunE == nil {
+			t.Error("PreRunE should be set to checkServerHeartbeat")
+		}
+	})
+}
+
+func TestRunIntegration_UnknownIntegration(t *testing.T) {
+	err := runIntegration("unknown-integration", "model")
+	if err == nil {
+		t.Error("expected error for unknown integration, got nil")
+	}
+	if !strings.Contains(err.Error(), "unknown integration") {
+		t.Errorf("error should mention 'unknown integration', got: %v", err)
+	}
+}
+
+func TestHasLocalModel_DocumentsHeuristic(t *testing.T) {
+	tests := []struct {
+		name   string
+		models []string
+		want   bool
+		reason string
+	}{
+		{"empty list", []string{}, false, "empty list has no local models"},
+		{"contains-cloud-substring", []string{"deepseek-r1:cloud"}, false, "model with 'cloud' substring is considered cloud"},
+		{"cloud-in-name", []string{"my-cloud-model"}, false, "'cloud' anywhere in name = cloud model"},
+		{"cloudless", []string{"cloudless-model"}, false, "'cloudless' still contains 'cloud'"},
+		{"local-model", []string{"llama3.2"}, true, "no 'cloud' = local"},
+		{"mixed", []string{"cloud-model", "llama3.2"}, true, "one local model = hasLocalModel true"},
+		{"all-cloud", []string{"cloud-a", "cloud-b"}, false, "all contain 'cloud'"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := slices.ContainsFunc(tt.models, func(m string) bool {
+				return !strings.Contains(m, "cloud")
+			})
+			if got != tt.want {
+				t.Errorf("hasLocalModel(%v) = %v, want %v (%s)", tt.models, got, tt.want, tt.reason)
+			}
+		})
+	}
+}
+
+func TestConfigCmd_NilHeartbeat(t *testing.T) {
+	// This should not panic - cmd creation should work even with nil
+	cmd := ConfigCmd(nil)
+	if cmd == nil {
+		t.Fatal("ConfigCmd returned nil")
+	}
+
+	// PreRunE should be nil when passed nil
+	if cmd.PreRunE != nil {
+		t.Log("Note: PreRunE is set even when nil is passed (acceptable)")
+	}
+}
+
+func TestAllIntegrations_HaveRequiredMethods(t *testing.T) {
+	for name, r := range integrations {
+		t.Run(name, func(t *testing.T) {
+			// Test String() doesn't panic and returns non-empty
+			displayName := r.String()
+			if displayName == "" {
+				t.Error("String() should not return empty")
+			}
+
+			// Test Run() exists (we can't call it without actually running the command)
+			// Just verify the method is available
+			var _ func(string) error = r.Run
+		})
+	}
+}

cmd/config/opencode.go

@@ -0,0 +1,203 @@
+package config
+
+import (
+	"encoding/json"
+	"fmt"
+	"maps"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"slices"
+	"strings"
+)
+
+// OpenCode implements Runner and Editor for OpenCode integration
+type OpenCode struct{}
+
+func (o *OpenCode) String() string { return "OpenCode" }
+
+func (o *OpenCode) Run(model string) error {
+	if _, err := exec.LookPath("opencode"); err != nil {
+		return fmt.Errorf("opencode is not installed, install from https://opencode.ai")
+	}
+
+	// Call Edit() to ensure config is up-to-date before launch
+	models := []string{model}
+	if config, err := loadIntegration("opencode"); err == nil && len(config.Models) > 0 {
+		models = config.Models
+	}
+	if err := o.Edit(models); err != nil {
+		return fmt.Errorf("setup failed: %w", err)
+	}
+
+	cmd := exec.Command("opencode")
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	return cmd.Run()
+}
+
+func (o *OpenCode) Paths() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+
+	var paths []string
+	p := filepath.Join(home, ".config", "opencode", "opencode.json")
+	if _, err := os.Stat(p); err == nil {
+		paths = append(paths, p)
+	}
+	sp := filepath.Join(home, ".local", "state", "opencode", "model.json")
+	if _, err := os.Stat(sp); err == nil {
+		paths = append(paths, sp)
+	}
+	return paths
+}
+
+func (o *OpenCode) Edit(modelList []string) error {
+	if len(modelList) == 0 {
+		return nil
+	}
+
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return err
+	}
+
+	configPath := filepath.Join(home, ".config", "opencode", "opencode.json")
+	if err := os.MkdirAll(filepath.Dir(configPath), 0o755); err != nil {
+		return err
+	}
+
+	config := make(map[string]any)
+	if data, err := os.ReadFile(configPath); err == nil {
+		_ = json.Unmarshal(data, &config) // Ignore parse errors; treat missing/corrupt files as empty
+	}
+
+	config["$schema"] = "https://opencode.ai/config.json"
+
+	provider, ok := config["provider"].(map[string]any)
+	if !ok {
+		provider = make(map[string]any)
+	}
+
+	ollama, ok := provider["ollama"].(map[string]any)
+	if !ok {
+		ollama = map[string]any{
+			"npm":  "@ai-sdk/openai-compatible",
+			"name": "Ollama (local)",
+			"options": map[string]any{
+				"baseURL": "http://localhost:11434/v1",
+			},
+		}
+	}
+
+	models, ok := ollama["models"].(map[string]any)
+	if !ok {
+		models = make(map[string]any)
+	}
+
+	selectedSet := make(map[string]bool)
+	for _, m := range modelList {
+		selectedSet[m] = true
+	}
+
+	for name, cfg := range models {
+		if cfgMap, ok := cfg.(map[string]any); ok {
+			if displayName, ok := cfgMap["name"].(string); ok {
+				if strings.HasSuffix(displayName, "[Ollama]") && !selectedSet[name] {
+					delete(models, name)
+				}
+			}
+		}
+	}
+
+	for _, model := range modelList {
+		models[model] = map[string]any{
+			"name": fmt.Sprintf("%s [Ollama]", model),
+		}
+	}
+
+	ollama["models"] = models
+	provider["ollama"] = ollama
+	config["provider"] = provider
+
+	configData, err := json.MarshalIndent(config, "", "  ")
+	if err != nil {
+		return err
+	}
+	if err := writeWithBackup(configPath, configData); err != nil {
+		return err
+	}
+
+	statePath := filepath.Join(home, ".local", "state", "opencode", "model.json")
+	if err := os.MkdirAll(filepath.Dir(statePath), 0o755); err != nil {
+		return err
+	}
+
+	state := map[string]any{
+		"recent":   []any{},
+		"favorite": []any{},
+		"variant":  map[string]any{},
+	}
+	if data, err := os.ReadFile(statePath); err == nil {
+		_ = json.Unmarshal(data, &state) // Ignore parse errors; use defaults
+	}
+
+	recent, _ := state["recent"].([]any)
+
+	modelSet := make(map[string]bool)
+	for _, m := range modelList {
+		modelSet[m] = true
+	}
+
+	// Filter out existing Ollama models we're about to re-add
+	newRecent := slices.DeleteFunc(slices.Clone(recent), func(entry any) bool {
+		e, ok := entry.(map[string]any)
+		if !ok || e["providerID"] != "ollama" {
+			return false
+		}
+		modelID, _ := e["modelID"].(string)
+		return modelSet[modelID]
+	})
+
+	// Prepend models in reverse order so first model ends up first
+	for _, model := range slices.Backward(modelList) {
+		newRecent = slices.Insert(newRecent, 0, any(map[string]any{
+			"providerID": "ollama",
+			"modelID":    model,
+		}))
+	}
+
+	const maxRecentModels = 10
+	newRecent = newRecent[:min(len(newRecent), maxRecentModels)]
+
+	state["recent"] = newRecent
+
+	stateData, err := json.MarshalIndent(state, "", "  ")
+	if err != nil {
+		return err
+	}
+	return writeWithBackup(statePath, stateData)
+}
+
+func (o *OpenCode) Models() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+	config, err := readJSONFile(filepath.Join(home, ".config", "opencode", "opencode.json"))
+	if err != nil {
+		return nil
+	}
+	provider, _ := config["provider"].(map[string]any)
+	ollama, _ := provider["ollama"].(map[string]any)
+	models, _ := ollama["models"].(map[string]any)
+	if len(models) == 0 {
+		return nil
+	}
+	keys := slices.Collect(maps.Keys(models))
+	slices.Sort(keys)
+	return keys
+}

cmd/config/opencode_test.go

@@ -0,0 +1,437 @@
+package config
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestOpenCodeIntegration(t *testing.T) {
+	o := &OpenCode{}
+
+	t.Run("String", func(t *testing.T) {
+		if got := o.String(); got != "OpenCode" {
+			t.Errorf("String() = %q, want %q", got, "OpenCode")
+		}
+	})
+
+	t.Run("implements Runner", func(t *testing.T) {
+		var _ Runner = o
+	})
+
+	t.Run("implements Editor", func(t *testing.T) {
+		var _ Editor = o
+	})
+}
+
+func TestOpenCodeEdit(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+	stateDir := filepath.Join(tmpDir, ".local", "state", "opencode")
+	statePath := filepath.Join(stateDir, "model.json")
+
+	cleanup := func() {
+		os.RemoveAll(configDir)
+		os.RemoveAll(stateDir)
+	}
+
+	t.Run("fresh install", func(t *testing.T) {
+		cleanup()
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+		assertOpenCodeRecentModel(t, statePath, 0, "ollama", "llama3.2")
+	})
+
+	t.Run("preserve other providers", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"provider":{"anthropic":{"apiKey":"xxx"}}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		provider := cfg["provider"].(map[string]any)
+		if provider["anthropic"] == nil {
+			t.Error("anthropic provider was removed")
+		}
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+	})
+
+	t.Run("preserve other models", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"provider":{"ollama":{"models":{"mistral":{"name":"Mistral"}}}}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenCodeModelExists(t, configPath, "mistral")
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+	})
+
+	t.Run("update existing model", func(t *testing.T) {
+		cleanup()
+		o.Edit([]string{"llama3.2"})
+		o.Edit([]string{"llama3.2"})
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+	})
+
+	t.Run("preserve top-level keys", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"theme":"dark","keybindings":{}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		if cfg["theme"] != "dark" {
+			t.Error("theme was removed")
+		}
+		if cfg["keybindings"] == nil {
+			t.Error("keybindings was removed")
+		}
+	})
+
+	t.Run("model state - insert at index 0", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(stateDir, 0o755)
+		os.WriteFile(statePath, []byte(`{"recent":[{"providerID":"anthropic","modelID":"claude"}],"favorite":[],"variant":{}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenCodeRecentModel(t, statePath, 0, "ollama", "llama3.2")
+		assertOpenCodeRecentModel(t, statePath, 1, "anthropic", "claude")
+	})
+
+	t.Run("model state - preserve favorites and variants", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(stateDir, 0o755)
+		os.WriteFile(statePath, []byte(`{"recent":[],"favorite":[{"providerID":"x","modelID":"y"}],"variant":{"a":"b"}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		data, _ := os.ReadFile(statePath)
+		var state map[string]any
+		json.Unmarshal(data, &state)
+		if len(state["favorite"].([]any)) != 1 {
+			t.Error("favorite was modified")
+		}
+		if state["variant"].(map[string]any)["a"] != "b" {
+			t.Error("variant was modified")
+		}
+	})
+
+	t.Run("model state - deduplicate on re-add", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(stateDir, 0o755)
+		os.WriteFile(statePath, []byte(`{"recent":[{"providerID":"ollama","modelID":"llama3.2"},{"providerID":"anthropic","modelID":"claude"}],"favorite":[],"variant":{}}`), 0o644)
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		data, _ := os.ReadFile(statePath)
+		var state map[string]any
+		json.Unmarshal(data, &state)
+		recent := state["recent"].([]any)
+		if len(recent) != 2 {
+			t.Errorf("expected 2 recent entries, got %d", len(recent))
+		}
+		assertOpenCodeRecentModel(t, statePath, 0, "ollama", "llama3.2")
+	})
+
+	t.Run("remove model", func(t *testing.T) {
+		cleanup()
+		// First add two models
+		o.Edit([]string{"llama3.2", "mistral"})
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+		assertOpenCodeModelExists(t, configPath, "mistral")
+
+		// Then remove one by only selecting the other
+		o.Edit([]string{"llama3.2"})
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+		assertOpenCodeModelNotExists(t, configPath, "mistral")
+	})
+
+	t.Run("remove model preserves non-ollama models", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		// Add a non-Ollama model manually
+		os.WriteFile(configPath, []byte(`{"provider":{"ollama":{"models":{"external":{"name":"External Model"}}}}}`), 0o644)
+
+		o.Edit([]string{"llama3.2"})
+		assertOpenCodeModelExists(t, configPath, "llama3.2")
+		assertOpenCodeModelExists(t, configPath, "external") // Should be preserved
+	})
+}
+
+func assertOpenCodeModelExists(t *testing.T, path, model string) {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var cfg map[string]any
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Fatal(err)
+	}
+	provider, ok := cfg["provider"].(map[string]any)
+	if !ok {
+		t.Fatal("provider not found")
+	}
+	ollama, ok := provider["ollama"].(map[string]any)
+	if !ok {
+		t.Fatal("ollama provider not found")
+	}
+	models, ok := ollama["models"].(map[string]any)
+	if !ok {
+		t.Fatal("models not found")
+	}
+	if models[model] == nil {
+		t.Errorf("model %s not found", model)
+	}
+}
+
+func assertOpenCodeModelNotExists(t *testing.T, path, model string) {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var cfg map[string]any
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Fatal(err)
+	}
+	provider, ok := cfg["provider"].(map[string]any)
+	if !ok {
+		return // No provider means no model
+	}
+	ollama, ok := provider["ollama"].(map[string]any)
+	if !ok {
+		return // No ollama means no model
+	}
+	models, ok := ollama["models"].(map[string]any)
+	if !ok {
+		return // No models means no model
+	}
+	if models[model] != nil {
+		t.Errorf("model %s should not exist but was found", model)
+	}
+}
+
+func assertOpenCodeRecentModel(t *testing.T, path string, index int, providerID, modelID string) {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var state map[string]any
+	if err := json.Unmarshal(data, &state); err != nil {
+		t.Fatal(err)
+	}
+	recent, ok := state["recent"].([]any)
+	if !ok {
+		t.Fatal("recent not found")
+	}
+	if index >= len(recent) {
+		t.Fatalf("index %d out of range (len=%d)", index, len(recent))
+	}
+	entry, ok := recent[index].(map[string]any)
+	if !ok {
+		t.Fatal("entry is not a map")
+	}
+	if entry["providerID"] != providerID {
+		t.Errorf("expected providerID %s, got %s", providerID, entry["providerID"])
+	}
+	if entry["modelID"] != modelID {
+		t.Errorf("expected modelID %s, got %s", modelID, entry["modelID"])
+	}
+}
+
+// Edge case tests for opencode.go
+
+func TestOpenCodeEdit_CorruptedConfigJSON(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(`{corrupted json content`), 0o644)
+
+	// Should not panic - corrupted JSON should be treated as empty
+	err := o.Edit([]string{"llama3.2"})
+	if err != nil {
+		t.Fatalf("Edit failed with corrupted config: %v", err)
+	}
+
+	// Verify valid JSON was created
+	data, _ := os.ReadFile(configPath)
+	var cfg map[string]any
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Errorf("resulting config is not valid JSON: %v", err)
+	}
+}
+
+func TestOpenCodeEdit_CorruptedStateJSON(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	stateDir := filepath.Join(tmpDir, ".local", "state", "opencode")
+	statePath := filepath.Join(stateDir, "model.json")
+
+	os.MkdirAll(stateDir, 0o755)
+	os.WriteFile(statePath, []byte(`{corrupted state`), 0o644)
+
+	err := o.Edit([]string{"llama3.2"})
+	if err != nil {
+		t.Fatalf("Edit failed with corrupted state: %v", err)
+	}
+
+	// Verify valid state was created
+	data, _ := os.ReadFile(statePath)
+	var state map[string]any
+	if err := json.Unmarshal(data, &state); err != nil {
+		t.Errorf("resulting state is not valid JSON: %v", err)
+	}
+}
+
+func TestOpenCodeEdit_WrongTypeProvider(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(`{"provider": "not a map"}`), 0o644)
+
+	err := o.Edit([]string{"llama3.2"})
+	if err != nil {
+		t.Fatalf("Edit with wrong type provider failed: %v", err)
+	}
+
+	// Verify provider is now correct type
+	data, _ := os.ReadFile(configPath)
+	var cfg map[string]any
+	json.Unmarshal(data, &cfg)
+
+	provider, ok := cfg["provider"].(map[string]any)
+	if !ok {
+		t.Fatalf("provider should be map after setup, got %T", cfg["provider"])
+	}
+	if provider["ollama"] == nil {
+		t.Error("ollama provider should be created")
+	}
+}
+
+func TestOpenCodeEdit_WrongTypeRecent(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	stateDir := filepath.Join(tmpDir, ".local", "state", "opencode")
+	statePath := filepath.Join(stateDir, "model.json")
+
+	os.MkdirAll(stateDir, 0o755)
+	os.WriteFile(statePath, []byte(`{"recent": "not an array", "favorite": [], "variant": {}}`), 0o644)
+
+	err := o.Edit([]string{"llama3.2"})
+	if err != nil {
+		t.Fatalf("Edit with wrong type recent failed: %v", err)
+	}
+
+	// The function should handle this gracefully
+	data, _ := os.ReadFile(statePath)
+	var state map[string]any
+	json.Unmarshal(data, &state)
+
+	// recent should be properly set after setup
+	recent, ok := state["recent"].([]any)
+	if !ok {
+		t.Logf("Note: recent type after setup is %T (documenting behavior)", state["recent"])
+	} else if len(recent) == 0 {
+		t.Logf("Note: recent is empty (documenting behavior)")
+	}
+}
+
+func TestOpenCodeEdit_EmptyModels(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+
+	os.MkdirAll(configDir, 0o755)
+	originalContent := `{"provider":{"ollama":{"models":{"existing":{}}}}}`
+	os.WriteFile(configPath, []byte(originalContent), 0o644)
+
+	// Empty models should be no-op
+	err := o.Edit([]string{})
+	if err != nil {
+		t.Fatalf("Edit with empty models failed: %v", err)
+	}
+
+	// Original content should be preserved (file not modified)
+	data, _ := os.ReadFile(configPath)
+	if string(data) != originalContent {
+		t.Errorf("empty models should not modify file, but content changed")
+	}
+}
+
+func TestOpenCodeEdit_SpecialCharsInModelName(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Model name with special characters (though unusual)
+	specialModel := `model-with-"quotes"`
+
+	err := o.Edit([]string{specialModel})
+	if err != nil {
+		t.Fatalf("Edit with special chars failed: %v", err)
+	}
+
+	// Verify it was stored correctly
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+	data, _ := os.ReadFile(configPath)
+
+	var cfg map[string]any
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Fatalf("resulting config is invalid JSON: %v", err)
+	}
+
+	// Model should be accessible
+	provider, _ := cfg["provider"].(map[string]any)
+	ollama, _ := provider["ollama"].(map[string]any)
+	models, _ := ollama["models"].(map[string]any)
+
+	if models[specialModel] == nil {
+		t.Errorf("model with special chars not found in config")
+	}
+}
+
+func TestOpenCodeModels_NoConfig(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	models := o.Models()
+	if len(models) > 0 {
+		t.Errorf("expected nil/empty for missing config, got %v", models)
+	}
+}

cmd/config/selector.go

@@ -0,0 +1,499 @@
+package config
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"strings"
+
+	"golang.org/x/term"
+)
+
+// ANSI escape sequences for terminal formatting.
+const (
+	ansiHideCursor = "\033[?25l"
+	ansiShowCursor = "\033[?25h"
+	ansiBold       = "\033[1m"
+	ansiReset      = "\033[0m"
+	ansiGray       = "\033[37m"
+	ansiClearDown  = "\033[J"
+)
+
+const maxDisplayedItems = 10
+
+var errCancelled = errors.New("cancelled")
+
+type selectItem struct {
+	Name        string
+	Description string
+}
+
+type inputEvent int
+
+const (
+	eventNone inputEvent = iota
+	eventEnter
+	eventEscape
+	eventUp
+	eventDown
+	eventTab
+	eventBackspace
+	eventChar
+)
+
+type selectState struct {
+	items        []selectItem
+	filter       string
+	selected     int
+	scrollOffset int
+}
+
+func newSelectState(items []selectItem) *selectState {
+	return &selectState{items: items}
+}
+
+func (s *selectState) filtered() []selectItem {
+	return filterItems(s.items, s.filter)
+}
+
+func (s *selectState) handleInput(event inputEvent, char byte) (done bool, result string, err error) {
+	filtered := s.filtered()
+
+	switch event {
+	case eventEnter:
+		if len(filtered) > 0 && s.selected < len(filtered) {
+			return true, filtered[s.selected].Name, nil
+		}
+	case eventEscape:
+		return true, "", errCancelled
+	case eventBackspace:
+		if len(s.filter) > 0 {
+			s.filter = s.filter[:len(s.filter)-1]
+			s.selected = 0
+			s.scrollOffset = 0
+		}
+	case eventUp:
+		if s.selected > 0 {
+			s.selected--
+			if s.selected < s.scrollOffset {
+				s.scrollOffset = s.selected
+			}
+		}
+	case eventDown:
+		if s.selected < len(filtered)-1 {
+			s.selected++
+			if s.selected >= s.scrollOffset+maxDisplayedItems {
+				s.scrollOffset = s.selected - maxDisplayedItems + 1
+			}
+		}
+	case eventChar:
+		s.filter += string(char)
+		s.selected = 0
+		s.scrollOffset = 0
+	}
+
+	return false, "", nil
+}
+
+type multiSelectState struct {
+	items         []selectItem
+	itemIndex     map[string]int
+	filter        string
+	highlighted   int
+	scrollOffset  int
+	checked       map[int]bool
+	checkOrder    []int
+	focusOnButton bool
+}
+
+func newMultiSelectState(items []selectItem, preChecked []string) *multiSelectState {
+	s := &multiSelectState{
+		items:     items,
+		itemIndex: make(map[string]int, len(items)),
+		checked:   make(map[int]bool),
+	}
+
+	for i, item := range items {
+		s.itemIndex[item.Name] = i
+	}
+
+	for _, name := range preChecked {
+		if idx, ok := s.itemIndex[name]; ok {
+			s.checked[idx] = true
+			s.checkOrder = append(s.checkOrder, idx)
+		}
+	}
+
+	return s
+}
+
+func (s *multiSelectState) filtered() []selectItem {
+	return filterItems(s.items, s.filter)
+}
+
+func (s *multiSelectState) toggleItem() {
+	filtered := s.filtered()
+	if len(filtered) == 0 || s.highlighted >= len(filtered) {
+		return
+	}
+
+	item := filtered[s.highlighted]
+	origIdx := s.itemIndex[item.Name]
+
+	if s.checked[origIdx] {
+		delete(s.checked, origIdx)
+		for i, idx := range s.checkOrder {
+			if idx == origIdx {
+				s.checkOrder = append(s.checkOrder[:i], s.checkOrder[i+1:]...)
+				break
+			}
+		}
+	} else {
+		s.checked[origIdx] = true
+		s.checkOrder = append(s.checkOrder, origIdx)
+	}
+}
+
+func (s *multiSelectState) handleInput(event inputEvent, char byte) (done bool, result []string, err error) {
+	filtered := s.filtered()
+
+	switch event {
+	case eventEnter:
+		if s.focusOnButton && len(s.checkOrder) > 0 {
+			var res []string
+			for _, idx := range s.checkOrder {
+				res = append(res, s.items[idx].Name)
+			}
+			return true, res, nil
+		} else if !s.focusOnButton {
+			s.toggleItem()
+		}
+	case eventTab:
+		if len(s.checkOrder) > 0 {
+			s.focusOnButton = !s.focusOnButton
+		}
+	case eventEscape:
+		return true, nil, errCancelled
+	case eventBackspace:
+		if len(s.filter) > 0 {
+			s.filter = s.filter[:len(s.filter)-1]
+			s.highlighted = 0
+			s.scrollOffset = 0
+			s.focusOnButton = false
+		}
+	case eventUp:
+		if s.focusOnButton {
+			s.focusOnButton = false
+		} else if s.highlighted > 0 {
+			s.highlighted--
+			if s.highlighted < s.scrollOffset {
+				s.scrollOffset = s.highlighted
+			}
+		}
+	case eventDown:
+		if s.focusOnButton {
+			s.focusOnButton = false
+		} else if s.highlighted < len(filtered)-1 {
+			s.highlighted++
+			if s.highlighted >= s.scrollOffset+maxDisplayedItems {
+				s.scrollOffset = s.highlighted - maxDisplayedItems + 1
+			}
+		}
+	case eventChar:
+		s.filter += string(char)
+		s.highlighted = 0
+		s.scrollOffset = 0
+		s.focusOnButton = false
+	}
+
+	return false, nil, nil
+}
+
+func (s *multiSelectState) selectedCount() int {
+	return len(s.checkOrder)
+}
+
+// Terminal I/O handling
+
+type terminalState struct {
+	fd       int
+	oldState *term.State
+}
+
+func enterRawMode() (*terminalState, error) {
+	fd := int(os.Stdin.Fd())
+	oldState, err := term.MakeRaw(fd)
+	if err != nil {
+		return nil, err
+	}
+	fmt.Fprint(os.Stderr, ansiHideCursor)
+	return &terminalState{fd: fd, oldState: oldState}, nil
+}
+
+func (t *terminalState) restore() {
+	fmt.Fprint(os.Stderr, ansiShowCursor)
+	term.Restore(t.fd, t.oldState)
+}
+
+func clearLines(n int) {
+	if n > 0 {
+		fmt.Fprintf(os.Stderr, "\033[%dA", n)
+		fmt.Fprint(os.Stderr, ansiClearDown)
+	}
+}
+
+func parseInput(r io.Reader) (inputEvent, byte, error) {
+	buf := make([]byte, 3)
+	n, err := r.Read(buf)
+	if err != nil {
+		return 0, 0, err
+	}
+
+	switch {
+	case n == 1 && buf[0] == 13:
+		return eventEnter, 0, nil
+	case n == 1 && (buf[0] == 3 || buf[0] == 27):
+		return eventEscape, 0, nil
+	case n == 1 && buf[0] == 9:
+		return eventTab, 0, nil
+	case n == 1 && buf[0] == 127:
+		return eventBackspace, 0, nil
+	case n == 3 && buf[0] == 27 && buf[1] == 91 && buf[2] == 65:
+		return eventUp, 0, nil
+	case n == 3 && buf[0] == 27 && buf[1] == 91 && buf[2] == 66:
+		return eventDown, 0, nil
+	case n == 1 && buf[0] >= 32 && buf[0] < 127:
+		return eventChar, buf[0], nil
+	}
+
+	return eventNone, 0, nil
+}
+
+// Rendering
+
+func renderSelect(w io.Writer, prompt string, s *selectState) int {
+	filtered := s.filtered()
+
+	fmt.Fprintf(w, "%s %s\r\n", prompt, s.filter)
+	lineCount := 1
+
+	if len(filtered) == 0 {
+		fmt.Fprintf(w, "  %s(no matches)%s\r\n", ansiGray, ansiReset)
+		lineCount++
+	} else {
+		displayCount := min(len(filtered), maxDisplayedItems)
+
+		for i := range displayCount {
+			idx := s.scrollOffset + i
+			if idx >= len(filtered) {
+				break
+			}
+			item := filtered[idx]
+			prefix := "    "
+			if idx == s.selected {
+				prefix = "  " + ansiBold + "> "
+			}
+			if item.Description != "" {
+				fmt.Fprintf(w, "%s%s%s %s- %s%s\r\n", prefix, item.Name, ansiReset, ansiGray, item.Description, ansiReset)
+			} else {
+				fmt.Fprintf(w, "%s%s%s\r\n", prefix, item.Name, ansiReset)
+			}
+			lineCount++
+		}
+
+		if remaining := len(filtered) - s.scrollOffset - displayCount; remaining > 0 {
+			fmt.Fprintf(w, "  %s... and %d more%s\r\n", ansiGray, remaining, ansiReset)
+			lineCount++
+		}
+	}
+
+	return lineCount
+}
+
+func renderMultiSelect(w io.Writer, prompt string, s *multiSelectState) int {
+	filtered := s.filtered()
+
+	fmt.Fprintf(w, "%s %s\r\n", prompt, s.filter)
+	lineCount := 1
+
+	if len(filtered) == 0 {
+		fmt.Fprintf(w, "  %s(no matches)%s\r\n", ansiGray, ansiReset)
+		lineCount++
+	} else {
+		displayCount := min(len(filtered), maxDisplayedItems)
+
+		for i := range displayCount {
+			idx := s.scrollOffset + i
+			if idx >= len(filtered) {
+				break
+			}
+			item := filtered[idx]
+			origIdx := s.itemIndex[item.Name]
+
+			checkbox := "[ ]"
+			if s.checked[origIdx] {
+				checkbox = "[x]"
+			}
+
+			prefix := "  "
+			suffix := ""
+			if idx == s.highlighted && !s.focusOnButton {
+				prefix = "> "
+			}
+			if len(s.checkOrder) > 0 && s.checkOrder[0] == origIdx {
+				suffix = " " + ansiGray + "(default)" + ansiReset
+			}
+
+			if idx == s.highlighted && !s.focusOnButton {
+				fmt.Fprintf(w, "  %s%s %s %s%s%s\r\n", ansiBold, prefix, checkbox, item.Name, ansiReset, suffix)
+			} else {
+				fmt.Fprintf(w, "  %s %s %s%s\r\n", prefix, checkbox, item.Name, suffix)
+			}
+			lineCount++
+		}
+
+		if remaining := len(filtered) - s.scrollOffset - displayCount; remaining > 0 {
+			fmt.Fprintf(w, "  %s... and %d more%s\r\n", ansiGray, remaining, ansiReset)
+			lineCount++
+		}
+	}
+
+	fmt.Fprintf(w, "\r\n")
+	lineCount++
+	count := s.selectedCount()
+	switch {
+	case count == 0:
+		fmt.Fprintf(w, "  %sSelect at least one model.%s\r\n", ansiGray, ansiReset)
+	case s.focusOnButton:
+		fmt.Fprintf(w, "  %s> [ Continue ]%s %s(%d selected)%s\r\n", ansiBold, ansiReset, ansiGray, count, ansiReset)
+	default:
+		fmt.Fprintf(w, "    %s[ Continue ] (%d selected) - press Tab%s\r\n", ansiGray, count, ansiReset)
+	}
+	lineCount++
+
+	return lineCount
+}
+
+// selectPrompt prompts the user to select a single item from a list.
+func selectPrompt(prompt string, items []selectItem) (string, error) {
+	if len(items) == 0 {
+		return "", fmt.Errorf("no items to select from")
+	}
+
+	ts, err := enterRawMode()
+	if err != nil {
+		return "", err
+	}
+	defer ts.restore()
+
+	state := newSelectState(items)
+	var lastLineCount int
+
+	render := func() {
+		clearLines(lastLineCount)
+		lastLineCount = renderSelect(os.Stderr, prompt, state)
+	}
+
+	render()
+
+	for {
+		event, char, err := parseInput(os.Stdin)
+		if err != nil {
+			return "", err
+		}
+
+		done, result, err := state.handleInput(event, char)
+		if done {
+			clearLines(lastLineCount)
+			if err != nil {
+				return "", err
+			}
+			return result, nil
+		}
+
+		render()
+	}
+}
+
+// multiSelectPrompt prompts the user to select multiple items from a list.
+func multiSelectPrompt(prompt string, items []selectItem, preChecked []string) ([]string, error) {
+	if len(items) == 0 {
+		return nil, fmt.Errorf("no items to select from")
+	}
+
+	ts, err := enterRawMode()
+	if err != nil {
+		return nil, err
+	}
+	defer ts.restore()
+
+	state := newMultiSelectState(items, preChecked)
+	var lastLineCount int
+
+	render := func() {
+		clearLines(lastLineCount)
+		lastLineCount = renderMultiSelect(os.Stderr, prompt, state)
+	}
+
+	render()
+
+	for {
+		event, char, err := parseInput(os.Stdin)
+		if err != nil {
+			return nil, err
+		}
+
+		done, result, err := state.handleInput(event, char)
+		if done {
+			clearLines(lastLineCount)
+			if err != nil {
+				return nil, err
+			}
+			return result, nil
+		}
+
+		render()
+	}
+}
+
+func confirmPrompt(prompt string) (bool, error) {
+	fd := int(os.Stdin.Fd())
+	oldState, err := term.MakeRaw(fd)
+	if err != nil {
+		return false, err
+	}
+	defer term.Restore(fd, oldState)
+
+	fmt.Fprintf(os.Stderr, "%s [y/n] ", prompt)
+
+	buf := make([]byte, 1)
+	for {
+		if _, err := os.Stdin.Read(buf); err != nil {
+			return false, err
+		}
+
+		switch buf[0] {
+		case 'Y', 'y', 13:
+			fmt.Fprintf(os.Stderr, "yes\r\n")
+			return true, nil
+		case 'N', 'n', 27, 3:
+			fmt.Fprintf(os.Stderr, "no\r\n")
+			return false, nil
+		}
+	}
+}
+
+func filterItems(items []selectItem, filter string) []selectItem {
+	if filter == "" {
+		return items
+	}
+	var result []selectItem
+	filterLower := strings.ToLower(filter)
+	for _, item := range items {
+		if strings.Contains(strings.ToLower(item.Name), filterLower) {
+			result = append(result, item)
+		}
+	}
+	return result
+}

cmd/config/selector_test.go

@@ -0,0 +1,913 @@
+package config
+
+import (
+	"bytes"
+	"strings"
+	"testing"
+)
+
+func TestFilterItems(t *testing.T) {
+	items := []selectItem{
+		{Name: "llama3.2:latest"},
+		{Name: "qwen2.5:7b"},
+		{Name: "deepseek-v3:cloud"},
+		{Name: "GPT-OSS:20b"},
+	}
+
+	t.Run("EmptyFilter_ReturnsAllItems", func(t *testing.T) {
+		result := filterItems(items, "")
+		if len(result) != len(items) {
+			t.Errorf("expected %d items, got %d", len(items), len(result))
+		}
+	})
+
+	t.Run("CaseInsensitive_UppercaseFilterMatchesLowercase", func(t *testing.T) {
+		result := filterItems(items, "LLAMA")
+		if len(result) != 1 || result[0].Name != "llama3.2:latest" {
+			t.Errorf("expected llama3.2:latest, got %v", result)
+		}
+	})
+
+	t.Run("CaseInsensitive_LowercaseFilterMatchesUppercase", func(t *testing.T) {
+		result := filterItems(items, "gpt")
+		if len(result) != 1 || result[0].Name != "GPT-OSS:20b" {
+			t.Errorf("expected GPT-OSS:20b, got %v", result)
+		}
+	})
+
+	t.Run("PartialMatch", func(t *testing.T) {
+		result := filterItems(items, "deep")
+		if len(result) != 1 || result[0].Name != "deepseek-v3:cloud" {
+			t.Errorf("expected deepseek-v3:cloud, got %v", result)
+		}
+	})
+
+	t.Run("NoMatch_ReturnsEmpty", func(t *testing.T) {
+		result := filterItems(items, "nonexistent")
+		if len(result) != 0 {
+			t.Errorf("expected 0 items, got %d", len(result))
+		}
+	})
+}
+
+func TestSelectState(t *testing.T) {
+	items := []selectItem{
+		{Name: "item1"},
+		{Name: "item2"},
+		{Name: "item3"},
+	}
+
+	t.Run("InitialState", func(t *testing.T) {
+		s := newSelectState(items)
+		if s.selected != 0 {
+			t.Errorf("expected selected=0, got %d", s.selected)
+		}
+		if s.filter != "" {
+			t.Errorf("expected empty filter, got %q", s.filter)
+		}
+		if s.scrollOffset != 0 {
+			t.Errorf("expected scrollOffset=0, got %d", s.scrollOffset)
+		}
+	})
+
+	t.Run("Enter_SelectsCurrentItem", func(t *testing.T) {
+		s := newSelectState(items)
+		done, result, err := s.handleInput(eventEnter, 0)
+		if !done || result != "item1" || err != nil {
+			t.Errorf("expected (true, item1, nil), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("Enter_WithFilter_SelectsFilteredItem", func(t *testing.T) {
+		s := newSelectState(items)
+		s.filter = "item3"
+		done, result, err := s.handleInput(eventEnter, 0)
+		if !done || result != "item3" || err != nil {
+			t.Errorf("expected (true, item3, nil), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("Enter_EmptyFilteredList_DoesNothing", func(t *testing.T) {
+		s := newSelectState(items)
+		s.filter = "nonexistent"
+		done, result, err := s.handleInput(eventEnter, 0)
+		if done || result != "" || err != nil {
+			t.Errorf("expected (false, '', nil), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("Escape_ReturnsCancelledError", func(t *testing.T) {
+		s := newSelectState(items)
+		done, result, err := s.handleInput(eventEscape, 0)
+		if !done || result != "" || err != errCancelled {
+			t.Errorf("expected (true, '', errCancelled), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("Down_MovesSelection", func(t *testing.T) {
+		s := newSelectState(items)
+		s.handleInput(eventDown, 0)
+		if s.selected != 1 {
+			t.Errorf("expected selected=1, got %d", s.selected)
+		}
+	})
+
+	t.Run("Down_AtBottom_StaysAtBottom", func(t *testing.T) {
+		s := newSelectState(items)
+		s.selected = 2
+		s.handleInput(eventDown, 0)
+		if s.selected != 2 {
+			t.Errorf("expected selected=2 (stayed at bottom), got %d", s.selected)
+		}
+	})
+
+	t.Run("Up_MovesSelection", func(t *testing.T) {
+		s := newSelectState(items)
+		s.selected = 2
+		s.handleInput(eventUp, 0)
+		if s.selected != 1 {
+			t.Errorf("expected selected=1, got %d", s.selected)
+		}
+	})
+
+	t.Run("Up_AtTop_StaysAtTop", func(t *testing.T) {
+		s := newSelectState(items)
+		s.handleInput(eventUp, 0)
+		if s.selected != 0 {
+			t.Errorf("expected selected=0 (stayed at top), got %d", s.selected)
+		}
+	})
+
+	t.Run("Char_AppendsToFilter", func(t *testing.T) {
+		s := newSelectState(items)
+		s.handleInput(eventChar, 'i')
+		s.handleInput(eventChar, 't')
+		s.handleInput(eventChar, 'e')
+		s.handleInput(eventChar, 'm')
+		s.handleInput(eventChar, '2')
+		if s.filter != "item2" {
+			t.Errorf("expected filter='item2', got %q", s.filter)
+		}
+		filtered := s.filtered()
+		if len(filtered) != 1 || filtered[0].Name != "item2" {
+			t.Errorf("expected [item2], got %v", filtered)
+		}
+	})
+
+	t.Run("Char_ResetsSelectionToZero", func(t *testing.T) {
+		s := newSelectState(items)
+		s.selected = 2
+		s.handleInput(eventChar, 'x')
+		if s.selected != 0 {
+			t.Errorf("expected selected=0 after typing, got %d", s.selected)
+		}
+	})
+
+	t.Run("Backspace_RemovesLastFilterChar", func(t *testing.T) {
+		s := newSelectState(items)
+		s.filter = "test"
+		s.handleInput(eventBackspace, 0)
+		if s.filter != "tes" {
+			t.Errorf("expected filter='tes', got %q", s.filter)
+		}
+	})
+
+	t.Run("Backspace_EmptyFilter_DoesNothing", func(t *testing.T) {
+		s := newSelectState(items)
+		s.handleInput(eventBackspace, 0)
+		if s.filter != "" {
+			t.Errorf("expected filter='', got %q", s.filter)
+		}
+	})
+
+	t.Run("Backspace_ResetsSelectionToZero", func(t *testing.T) {
+		s := newSelectState(items)
+		s.filter = "test"
+		s.selected = 2
+		s.handleInput(eventBackspace, 0)
+		if s.selected != 0 {
+			t.Errorf("expected selected=0 after backspace, got %d", s.selected)
+		}
+	})
+
+	t.Run("Scroll_DownPastVisibleItems_ScrollsViewport", func(t *testing.T) {
+		// maxDisplayedItems is 10, so with 15 items we need to scroll
+		manyItems := make([]selectItem, 15)
+		for i := range manyItems {
+			manyItems[i] = selectItem{Name: string(rune('a' + i))}
+		}
+		s := newSelectState(manyItems)
+
+		// move down 12 times (past the 10-item viewport)
+		for range 12 {
+			s.handleInput(eventDown, 0)
+		}
+
+		if s.selected != 12 {
+			t.Errorf("expected selected=12, got %d", s.selected)
+		}
+		if s.scrollOffset != 3 {
+			t.Errorf("expected scrollOffset=3 (12-10+1), got %d", s.scrollOffset)
+		}
+	})
+
+	t.Run("Scroll_UpPastScrollOffset_ScrollsViewport", func(t *testing.T) {
+		manyItems := make([]selectItem, 15)
+		for i := range manyItems {
+			manyItems[i] = selectItem{Name: string(rune('a' + i))}
+		}
+		s := newSelectState(manyItems)
+		s.selected = 5
+		s.scrollOffset = 5
+
+		s.handleInput(eventUp, 0)
+
+		if s.selected != 4 {
+			t.Errorf("expected selected=4, got %d", s.selected)
+		}
+		if s.scrollOffset != 4 {
+			t.Errorf("expected scrollOffset=4, got %d", s.scrollOffset)
+		}
+	})
+}
+
+func TestMultiSelectState(t *testing.T) {
+	items := []selectItem{
+		{Name: "item1"},
+		{Name: "item2"},
+		{Name: "item3"},
+	}
+
+	t.Run("InitialState_NoPrechecked", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		if s.highlighted != 0 {
+			t.Errorf("expected highlighted=0, got %d", s.highlighted)
+		}
+		if s.selectedCount() != 0 {
+			t.Errorf("expected 0 selected, got %d", s.selectedCount())
+		}
+		if s.focusOnButton {
+			t.Error("expected focusOnButton=false initially")
+		}
+	})
+
+	t.Run("InitialState_WithPrechecked", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item2", "item3"})
+		if s.selectedCount() != 2 {
+			t.Errorf("expected 2 selected, got %d", s.selectedCount())
+		}
+		if !s.checked[1] || !s.checked[2] {
+			t.Error("expected item2 and item3 to be checked")
+		}
+	})
+
+	t.Run("Prechecked_PreservesSelectionOrder", func(t *testing.T) {
+		// order matters: first checked = default model
+		s := newMultiSelectState(items, []string{"item3", "item1"})
+		if len(s.checkOrder) != 2 {
+			t.Fatalf("expected 2 in checkOrder, got %d", len(s.checkOrder))
+		}
+		if s.checkOrder[0] != 2 || s.checkOrder[1] != 0 {
+			t.Errorf("expected checkOrder=[2,0] (item3 first), got %v", s.checkOrder)
+		}
+	})
+
+	t.Run("Prechecked_IgnoresInvalidNames", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1", "nonexistent"})
+		if s.selectedCount() != 1 {
+			t.Errorf("expected 1 selected (nonexistent ignored), got %d", s.selectedCount())
+		}
+	})
+
+	t.Run("Toggle_ChecksUncheckedItem", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.toggleItem()
+		if !s.checked[0] {
+			t.Error("expected item1 to be checked after toggle")
+		}
+	})
+
+	t.Run("Toggle_UnchecksCheckedItem", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		s.toggleItem()
+		if s.checked[0] {
+			t.Error("expected item1 to be unchecked after toggle")
+		}
+	})
+
+	t.Run("Toggle_RemovesFromCheckOrder", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1", "item2", "item3"})
+		s.highlighted = 1 // toggle item2
+		s.toggleItem()
+
+		if len(s.checkOrder) != 2 {
+			t.Fatalf("expected 2 in checkOrder, got %d", len(s.checkOrder))
+		}
+		// should be [0, 2] (item1, item3) with item2 removed
+		if s.checkOrder[0] != 0 || s.checkOrder[1] != 2 {
+			t.Errorf("expected checkOrder=[0,2], got %v", s.checkOrder)
+		}
+	})
+
+	t.Run("Enter_TogglesWhenNotOnButton", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.handleInput(eventEnter, 0)
+		if !s.checked[0] {
+			t.Error("expected item1 to be checked after enter")
+		}
+	})
+
+	t.Run("Enter_OnButton_ReturnsSelection", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item2", "item1"})
+		s.focusOnButton = true
+
+		done, result, err := s.handleInput(eventEnter, 0)
+
+		if !done || err != nil {
+			t.Errorf("expected done=true, err=nil, got done=%v, err=%v", done, err)
+		}
+		// result should preserve selection order
+		if len(result) != 2 || result[0] != "item2" || result[1] != "item1" {
+			t.Errorf("expected [item2, item1], got %v", result)
+		}
+	})
+
+	t.Run("Enter_OnButton_EmptySelection_DoesNothing", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.focusOnButton = true
+		done, result, err := s.handleInput(eventEnter, 0)
+		if done || result != nil || err != nil {
+			t.Errorf("expected (false, nil, nil), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("Tab_SwitchesToButton_WhenHasSelection", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		s.handleInput(eventTab, 0)
+		if !s.focusOnButton {
+			t.Error("expected focus on button after tab")
+		}
+	})
+
+	t.Run("Tab_DoesNothing_WhenNoSelection", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.handleInput(eventTab, 0)
+		if s.focusOnButton {
+			t.Error("tab should not focus button when nothing selected")
+		}
+	})
+
+	t.Run("Tab_TogglesButtonFocus", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		s.handleInput(eventTab, 0)
+		if !s.focusOnButton {
+			t.Error("expected focus on button after first tab")
+		}
+		s.handleInput(eventTab, 0)
+		if s.focusOnButton {
+			t.Error("expected focus back on list after second tab")
+		}
+	})
+
+	t.Run("Escape_ReturnsCancelledError", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		done, result, err := s.handleInput(eventEscape, 0)
+		if !done || result != nil || err != errCancelled {
+			t.Errorf("expected (true, nil, errCancelled), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
+	t.Run("IsDefault_TrueForFirstChecked", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item2", "item1"})
+		if !(len(s.checkOrder) > 0 && s.checkOrder[0] == 1) {
+			t.Error("expected item2 (idx 1) to be default (first checked)")
+		}
+		if len(s.checkOrder) > 0 && s.checkOrder[0] == 0 {
+			t.Error("expected item1 (idx 0) to NOT be default")
+		}
+	})
+
+	t.Run("IsDefault_FalseWhenNothingChecked", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		if len(s.checkOrder) > 0 && s.checkOrder[0] == 0 {
+			t.Error("expected isDefault=false when nothing checked")
+		}
+	})
+
+	t.Run("Down_MovesHighlight", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.handleInput(eventDown, 0)
+		if s.highlighted != 1 {
+			t.Errorf("expected highlighted=1, got %d", s.highlighted)
+		}
+	})
+
+	t.Run("Up_MovesHighlight", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.highlighted = 1
+		s.handleInput(eventUp, 0)
+		if s.highlighted != 0 {
+			t.Errorf("expected highlighted=0, got %d", s.highlighted)
+		}
+	})
+
+	t.Run("Arrow_ReturnsFocusFromButton", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		s.focusOnButton = true
+		s.handleInput(eventDown, 0)
+		if s.focusOnButton {
+			t.Error("expected focus to return to list on arrow key")
+		}
+	})
+
+	t.Run("Char_AppendsToFilter", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.handleInput(eventChar, 'x')
+		if s.filter != "x" {
+			t.Errorf("expected filter='x', got %q", s.filter)
+		}
+	})
+
+	t.Run("Char_ResetsHighlightAndScroll", func(t *testing.T) {
+		manyItems := make([]selectItem, 15)
+		for i := range manyItems {
+			manyItems[i] = selectItem{Name: string(rune('a' + i))}
+		}
+		s := newMultiSelectState(manyItems, nil)
+		s.highlighted = 10
+		s.scrollOffset = 5
+
+		s.handleInput(eventChar, 'x')
+
+		if s.highlighted != 0 {
+			t.Errorf("expected highlighted=0, got %d", s.highlighted)
+		}
+		if s.scrollOffset != 0 {
+			t.Errorf("expected scrollOffset=0, got %d", s.scrollOffset)
+		}
+	})
+
+	t.Run("Backspace_RemovesLastFilterChar", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		s.filter = "test"
+		s.handleInput(eventBackspace, 0)
+		if s.filter != "tes" {
+			t.Errorf("expected filter='tes', got %q", s.filter)
+		}
+	})
+
+	t.Run("Backspace_RemovesFocusFromButton", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		s.filter = "x"
+		s.focusOnButton = true
+		s.handleInput(eventBackspace, 0)
+		if s.focusOnButton {
+			t.Error("expected focusOnButton=false after backspace")
+		}
+	})
+}
+
+func TestParseInput(t *testing.T) {
+	t.Run("Enter", func(t *testing.T) {
+		event, char, err := parseInput(bytes.NewReader([]byte{13}))
+		if err != nil || event != eventEnter || char != 0 {
+			t.Errorf("expected (eventEnter, 0, nil), got (%v, %v, %v)", event, char, err)
+		}
+	})
+
+	t.Run("Escape", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{27}))
+		if err != nil || event != eventEscape {
+			t.Errorf("expected eventEscape, got %v", event)
+		}
+	})
+
+	t.Run("CtrlC_TreatedAsEscape", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{3}))
+		if err != nil || event != eventEscape {
+			t.Errorf("expected eventEscape for Ctrl+C, got %v", event)
+		}
+	})
+
+	t.Run("Tab", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{9}))
+		if err != nil || event != eventTab {
+			t.Errorf("expected eventTab, got %v", event)
+		}
+	})
+
+	t.Run("Backspace", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{127}))
+		if err != nil || event != eventBackspace {
+			t.Errorf("expected eventBackspace, got %v", event)
+		}
+	})
+
+	t.Run("UpArrow", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{27, 91, 65}))
+		if err != nil || event != eventUp {
+			t.Errorf("expected eventUp, got %v", event)
+		}
+	})
+
+	t.Run("DownArrow", func(t *testing.T) {
+		event, _, err := parseInput(bytes.NewReader([]byte{27, 91, 66}))
+		if err != nil || event != eventDown {
+			t.Errorf("expected eventDown, got %v", event)
+		}
+	})
+
+	t.Run("PrintableChars", func(t *testing.T) {
+		tests := []struct {
+			name string
+			char byte
+		}{
+			{"lowercase", 'a'},
+			{"uppercase", 'Z'},
+			{"digit", '5'},
+			{"space", ' '},
+			{"tilde", '~'},
+		}
+		for _, tt := range tests {
+			t.Run(tt.name, func(t *testing.T) {
+				event, char, err := parseInput(bytes.NewReader([]byte{tt.char}))
+				if err != nil || event != eventChar || char != tt.char {
+					t.Errorf("expected (eventChar, %q), got (%v, %q)", tt.char, event, char)
+				}
+			})
+		}
+	})
+}
+
+func TestRenderSelect(t *testing.T) {
+	items := []selectItem{
+		{Name: "item1", Description: "first item"},
+		{Name: "item2"},
+	}
+
+	t.Run("ShowsPromptAndItems", func(t *testing.T) {
+		s := newSelectState(items)
+		var buf bytes.Buffer
+		lineCount := renderSelect(&buf, "Select:", s)
+
+		output := buf.String()
+		if !strings.Contains(output, "Select:") {
+			t.Error("expected prompt in output")
+		}
+		if !strings.Contains(output, "item1") {
+			t.Error("expected item1 in output")
+		}
+		if !strings.Contains(output, "first item") {
+			t.Error("expected description in output")
+		}
+		if !strings.Contains(output, "item2") {
+			t.Error("expected item2 in output")
+		}
+		if lineCount != 3 { // 1 prompt + 2 items
+			t.Errorf("expected 3 lines, got %d", lineCount)
+		}
+	})
+
+	t.Run("EmptyFilteredList_ShowsNoMatches", func(t *testing.T) {
+		s := newSelectState(items)
+		s.filter = "xyz"
+		var buf bytes.Buffer
+		renderSelect(&buf, "Select:", s)
+
+		if !strings.Contains(buf.String(), "no matches") {
+			t.Error("expected 'no matches' message")
+		}
+	})
+
+	t.Run("LongList_ShowsRemainingCount", func(t *testing.T) {
+		manyItems := make([]selectItem, 15)
+		for i := range manyItems {
+			manyItems[i] = selectItem{Name: string(rune('a' + i))}
+		}
+		s := newSelectState(manyItems)
+		var buf bytes.Buffer
+		renderSelect(&buf, "Select:", s)
+
+		// 15 items - 10 displayed = 5 more
+		if !strings.Contains(buf.String(), "5 more") {
+			t.Error("expected '5 more' indicator")
+		}
+	})
+}
+
+func TestRenderMultiSelect(t *testing.T) {
+	items := []selectItem{
+		{Name: "item1"},
+		{Name: "item2"},
+	}
+
+	t.Run("ShowsCheckboxes", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		var buf bytes.Buffer
+		renderMultiSelect(&buf, "Select:", s)
+
+		output := buf.String()
+		if !strings.Contains(output, "[x]") {
+			t.Error("expected checked checkbox [x]")
+		}
+		if !strings.Contains(output, "[ ]") {
+			t.Error("expected unchecked checkbox [ ]")
+		}
+	})
+
+	t.Run("ShowsDefaultMarker", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1"})
+		var buf bytes.Buffer
+		renderMultiSelect(&buf, "Select:", s)
+
+		if !strings.Contains(buf.String(), "(default)") {
+			t.Error("expected (default) marker for first checked item")
+		}
+	})
+
+	t.Run("ShowsSelectedCount", func(t *testing.T) {
+		s := newMultiSelectState(items, []string{"item1", "item2"})
+		var buf bytes.Buffer
+		renderMultiSelect(&buf, "Select:", s)
+
+		if !strings.Contains(buf.String(), "2 selected") {
+			t.Error("expected '2 selected' in output")
+		}
+	})
+
+	t.Run("NoSelection_ShowsHelperText", func(t *testing.T) {
+		s := newMultiSelectState(items, nil)
+		var buf bytes.Buffer
+		renderMultiSelect(&buf, "Select:", s)
+
+		if !strings.Contains(buf.String(), "Select at least one") {
+			t.Error("expected 'Select at least one' helper text")
+		}
+	})
+}
+
+func TestErrCancelled(t *testing.T) {
+	t.Run("NotNil", func(t *testing.T) {
+		if errCancelled == nil {
+			t.Error("errCancelled should not be nil")
+		}
+	})
+
+	t.Run("Message", func(t *testing.T) {
+		if errCancelled.Error() != "cancelled" {
+			t.Errorf("expected 'cancelled', got %q", errCancelled.Error())
+		}
+	})
+}
+
+// Edge case tests for selector.go
+
+// TestSelectState_SingleItem verifies that single item list works without crash.
+// List with only one item should still work.
+func TestSelectState_SingleItem(t *testing.T) {
+	items := []selectItem{{Name: "only-one"}}
+
+	s := newSelectState(items)
+
+	// Down should do nothing (already at bottom)
+	s.handleInput(eventDown, 0)
+	if s.selected != 0 {
+		t.Errorf("down on single item: expected selected=0, got %d", s.selected)
+	}
+
+	// Up should do nothing (already at top)
+	s.handleInput(eventUp, 0)
+	if s.selected != 0 {
+		t.Errorf("up on single item: expected selected=0, got %d", s.selected)
+	}
+
+	// Enter should select the only item
+	done, result, err := s.handleInput(eventEnter, 0)
+	if !done || result != "only-one" || err != nil {
+		t.Errorf("enter on single item: expected (true, 'only-one', nil), got (%v, %q, %v)", done, result, err)
+	}
+}
+
+// TestSelectState_ExactlyMaxItems verifies boundary condition at maxDisplayedItems.
+// List with exactly maxDisplayedItems items should not scroll.
+func TestSelectState_ExactlyMaxItems(t *testing.T) {
+	items := make([]selectItem, maxDisplayedItems)
+	for i := range items {
+		items[i] = selectItem{Name: string(rune('a' + i))}
+	}
+
+	s := newSelectState(items)
+
+	// Move to last item
+	for range maxDisplayedItems - 1 {
+		s.handleInput(eventDown, 0)
+	}
+
+	if s.selected != maxDisplayedItems-1 {
+		t.Errorf("expected selected=%d, got %d", maxDisplayedItems-1, s.selected)
+	}
+
+	// Should not scroll when exactly at max
+	if s.scrollOffset != 0 {
+		t.Errorf("expected scrollOffset=0 for exactly maxDisplayedItems, got %d", s.scrollOffset)
+	}
+
+	// One more down should do nothing
+	s.handleInput(eventDown, 0)
+	if s.selected != maxDisplayedItems-1 {
+		t.Errorf("down at max: expected selected=%d, got %d", maxDisplayedItems-1, s.selected)
+	}
+}
+
+// TestFilterItems_RegexSpecialChars verifies that filter is literal, not regex.
+// User typing "model.v1" shouldn't match "modelsv1".
+func TestFilterItems_RegexSpecialChars(t *testing.T) {
+	items := []selectItem{
+		{Name: "model.v1"},
+		{Name: "modelsv1"},
+		{Name: "model-v1"},
+	}
+
+	// Filter with dot should only match literal dot
+	result := filterItems(items, "model.v1")
+	if len(result) != 1 {
+		t.Errorf("expected 1 exact match, got %d", len(result))
+	}
+	if len(result) > 0 && result[0].Name != "model.v1" {
+		t.Errorf("expected 'model.v1', got %s", result[0].Name)
+	}
+
+	// Other regex special chars should be literal too
+	items2 := []selectItem{
+		{Name: "test[0]"},
+		{Name: "test0"},
+		{Name: "test(1)"},
+	}
+
+	result2 := filterItems(items2, "test[0]")
+	if len(result2) != 1 || result2[0].Name != "test[0]" {
+		t.Errorf("expected only 'test[0]', got %v", result2)
+	}
+}
+
+// TestMultiSelectState_DuplicateNames documents handling of duplicate item names.
+// itemIndex uses name as key - duplicates cause collision. This documents
+// the current behavior: the last index for a duplicate name is stored
+func TestMultiSelectState_DuplicateNames(t *testing.T) {
+	// Duplicate names - this is an edge case that shouldn't happen in practice
+	items := []selectItem{
+		{Name: "duplicate"},
+		{Name: "duplicate"},
+		{Name: "unique"},
+	}
+
+	s := newMultiSelectState(items, nil)
+
+	// DOCUMENTED BEHAVIOR: itemIndex maps name to LAST index
+	// When there are duplicates, only the last occurrence's index is stored
+	if s.itemIndex["duplicate"] != 1 {
+		t.Errorf("itemIndex should map 'duplicate' to last index (1), got %d", s.itemIndex["duplicate"])
+	}
+
+	// Toggle item at highlighted=0 (first "duplicate")
+	// Due to name collision, toggleItem uses itemIndex["duplicate"] = 1
+	// So it actually toggles the SECOND duplicate item, not the first
+	s.toggleItem()
+
+	// This documents the potentially surprising behavior:
+	// We toggled at highlighted=0, but itemIndex lookup returned 1
+	if !s.checked[1] {
+		t.Error("toggle should check index 1 (due to name collision in itemIndex)")
+	}
+	if s.checked[0] {
+		t.Log("Note: index 0 is NOT checked, even though highlighted=0 (name collision behavior)")
+	}
+}
+
+// TestSelectState_FilterReducesBelowSelection verifies selection resets when filter reduces list.
+// Prevents index-out-of-bounds on next keystroke
+func TestSelectState_FilterReducesBelowSelection(t *testing.T) {
+	items := []selectItem{
+		{Name: "apple"},
+		{Name: "banana"},
+		{Name: "cherry"},
+	}
+
+	s := newSelectState(items)
+	s.selected = 2 // Select "cherry"
+
+	// Type a filter that removes cherry from results
+	s.handleInput(eventChar, 'a') // Filter to "a" - matches "apple" and "banana"
+
+	// Selection should reset to 0
+	if s.selected != 0 {
+		t.Errorf("expected selected=0 after filter, got %d", s.selected)
+	}
+
+	filtered := s.filtered()
+	if len(filtered) != 2 {
+		t.Errorf("expected 2 filtered items, got %d", len(filtered))
+	}
+}
+
+// TestFilterItems_UnicodeCharacters verifies filtering works with UTF-8.
+// Model names might contain unicode characters
+func TestFilterItems_UnicodeCharacters(t *testing.T) {
+	items := []selectItem{
+		{Name: "llama-日本語"},
+		{Name: "模型-chinese"},
+		{Name: "émoji-🦙"},
+		{Name: "regular-model"},
+	}
+
+	t.Run("filter japanese", func(t *testing.T) {
+		result := filterItems(items, "日本")
+		if len(result) != 1 || result[0].Name != "llama-日本語" {
+			t.Errorf("expected llama-日本語, got %v", result)
+		}
+	})
+
+	t.Run("filter chinese", func(t *testing.T) {
+		result := filterItems(items, "模型")
+		if len(result) != 1 || result[0].Name != "模型-chinese" {
+			t.Errorf("expected 模型-chinese, got %v", result)
+		}
+	})
+
+	t.Run("filter emoji", func(t *testing.T) {
+		result := filterItems(items, "🦙")
+		if len(result) != 1 || result[0].Name != "émoji-🦙" {
+			t.Errorf("expected émoji-🦙, got %v", result)
+		}
+	})
+
+	t.Run("filter accented char", func(t *testing.T) {
+		result := filterItems(items, "émoji")
+		if len(result) != 1 || result[0].Name != "émoji-🦙" {
+			t.Errorf("expected émoji-🦙, got %v", result)
+		}
+	})
+}
+
+// TestMultiSelectState_FilterReducesBelowHighlight verifies highlight resets when filter reduces list.
+func TestMultiSelectState_FilterReducesBelowHighlight(t *testing.T) {
+	items := []selectItem{
+		{Name: "apple"},
+		{Name: "banana"},
+		{Name: "cherry"},
+	}
+
+	s := newMultiSelectState(items, nil)
+	s.highlighted = 2 // Highlight "cherry"
+
+	// Type a filter that removes cherry
+	s.handleInput(eventChar, 'a')
+
+	if s.highlighted != 0 {
+		t.Errorf("expected highlighted=0 after filter, got %d", s.highlighted)
+	}
+}
+
+// TestMultiSelectState_EmptyItems verifies handling of empty item list.
+// Empty list should be handled gracefully.
+func TestMultiSelectState_EmptyItems(t *testing.T) {
+	s := newMultiSelectState([]selectItem{}, nil)
+
+	// Toggle should not panic on empty list
+	s.toggleItem()
+
+	if s.selectedCount() != 0 {
+		t.Errorf("expected 0 selected for empty list, got %d", s.selectedCount())
+	}
+
+	// Render should handle empty list
+	var buf bytes.Buffer
+	lineCount := renderMultiSelect(&buf, "Select:", s)
+	if lineCount == 0 {
+		t.Error("renderMultiSelect should produce output even for empty list")
+	}
+	if !strings.Contains(buf.String(), "no matches") {
+		t.Error("expected 'no matches' for empty list")
+	}
+}
+
+// TestSelectState_RenderWithDescriptions verifies rendering items with descriptions.
+func TestSelectState_RenderWithDescriptions(t *testing.T) {
+	items := []selectItem{
+		{Name: "item1", Description: "First item description"},
+		{Name: "item2", Description: ""},
+		{Name: "item3", Description: "Third item"},
+	}
+
+	s := newSelectState(items)
+	var buf bytes.Buffer
+	renderSelect(&buf, "Select:", s)
+
+	output := buf.String()
+	if !strings.Contains(output, "First item description") {
+		t.Error("expected description to be rendered")
+	}
+	if !strings.Contains(output, "item2") {
+		t.Error("expected item without description to be rendered")
+	}
+}

cmd/interactive.go

@@ -116,7 +116,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		Prompt:         ">>> ",
 		AltPrompt:      "... ",
 		Placeholder:    "Send a message (/? for help)",
-		AltPlaceholder: `Use """ to end multi-line input`,
+		AltPlaceholder: "Press Enter to send",
 	})
 	if err != nil {
 		return err
@@ -159,6 +159,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			sb.WriteString(before)
 			if !ok {
 				fmt.Fprintln(&sb)
+				scanner.Prompt.UseAlt = true
 				continue
 			}
 

convert/convert.go

@@ -6,11 +6,14 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
+	"iter"
 	"log/slog"
+	"maps"
 	"os"
 	"slices"
 	"strings"
 
+	ofs "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -18,8 +21,13 @@ type ModelParameters struct {
 	Architectures []string `json:"architectures"`
 	VocabSize     uint32   `json:"vocab_size"`
 
+	// TODO is this needed?
+	ModelType string `json:"model_type"`
+
 	TextModel struct {
-		VocabSize uint32 `json:"vocab_size"`
+		VocabSize  uint32 `json:"vocab_size"`
+		HiddenSize uint32 `json:"hidden_size"`
+		ModelType  string `json:"model_type"`
 	} `json:"text_config"`
 }
 
@@ -33,8 +41,94 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }
 
-func (ModelParameters) KV(t *Tokenizer) ggml.KV {
-	kv := ggml.KV{
+type KV map[string]any
+
+func (kv KV) Architecture() string {
+	return kv.String("general.architecture", "unknown")
+}
+
+type valueTypes interface {
+	uint8 | int8 | uint16 | int16 |
+		uint32 | int32 | uint64 | int64 |
+		string | float32 | float64 | bool
+}
+
+type arrayValueTypes interface {
+	[]uint8 | []int8 | []uint16 | []int16 |
+		[]uint32 | []int32 | []uint64 | []int64 |
+		[]string | []float32 | []float64 | []bool
+}
+
+func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ...T) (T, bool) {
+	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
+		key = kv.Architecture() + "." + key
+	}
+
+	if val, ok := kv[key].(T); ok {
+		return val, true
+	}
+	return defaultValue[0], false
+}
+
+func (kv KV) String(key string, defaultValue ...string) string {
+	val, _ := keyValue(kv, key, append(defaultValue, "")...)
+	return val
+}
+
+func (kv KV) Uint(key string, defaultValue ...uint32) uint32 {
+	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
+	return val
+}
+
+func (kv KV) Float(key string, defaultValue ...float32) float32 {
+	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
+	return val
+}
+
+func (kv KV) Bool(key string, defaultValue ...bool) bool {
+	val, _ := keyValue(kv, key, append(defaultValue, false)...)
+	return val
+}
+
+func (kv KV) Strings(key string, defaultValue ...[]string) []string {
+	val, _ := keyValue(kv, key, append(defaultValue, []string{""})...)
+	return val
+}
+
+func (kv KV) Ints(key string, defaultValue ...[]int32) []int32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []int32{0})...)
+	return val
+}
+
+func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []uint32{0})...)
+	return val
+}
+
+func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []float32{0})...)
+	return val
+}
+
+func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
+	val, _ := keyValue(kv, key, append(defaultValue, []bool{false})...)
+	return val
+}
+
+func (kv KV) Len() int {
+	return len(kv)
+}
+
+func (kv KV) Keys() iter.Seq[string] {
+	return maps.Keys(kv)
+}
+
+func (kv KV) Value(key string) any {
+	return kv[key]
+}
+
+func (ModelParameters) KV(t *Tokenizer) KV {
+	kv := KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -63,7 +157,7 @@ func (ModelParameters) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p AdapterParameters) KV() ggml.KV {
+func (p AdapterParameters) KV() KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -71,7 +165,7 @@ func (p AdapterParameters) KV() ggml.KV {
 		alpha = p.LoraParameters.Alpha
 	}
 
-	kv := ggml.KV{
+	kv := KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -88,9 +182,14 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-type ModelConverter interface {
+type ModelKV interface {
 	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) ggml.KV
+	KV(*Tokenizer) KV
+}
+
+type ModelConverter interface {
+	ModelKV
+
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
 	Tensors([]Tensor) []*ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
@@ -107,7 +206,7 @@ type moreParser interface {
 
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(ggml.KV) ggml.KV
+	KV(ofs.Config) KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
 	Tensors([]Tensor) []*ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
@@ -115,7 +214,7 @@ type AdapterConverter interface {
 	Replacements() []string
 }
 
-func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
+func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ofs.Config) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err
@@ -126,8 +225,8 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 		return err
 	}
 
-	arch, ok := baseKV["general.architecture"]
-	if !ok {
+	arch := baseKV.Architecture()
+	if arch == "" {
 		return errors.New("architecture not set for the base model")
 	}
 
@@ -153,23 +252,19 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 	return writeFile(f, conv.KV(baseKV), conv.Tensors(ts))
 }
 
-// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
-// and files it finds in the input path.
-// Supported input model formats include safetensors.
-// Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
-func ConvertModel(fsys fs.FS, f *os.File) error {
+func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 	bts, err := fs.ReadFile(fsys, "config.json")
 	if err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	var p ModelParameters
 	if err := json.Unmarshal(bts, &p); err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	if len(p.Architectures) < 1 {
-		return errors.New("unknown architecture")
+		return nil, nil, errors.New("unknown architecture")
 	}
 
 	var conv ModelConverter
@@ -216,23 +311,27 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &deepseekocr{}
 	case "DeepseekV3ForCausalLM":
 		conv = &deepseek2Model{}
+	case "Glm4MoeLiteForCausalLM":
+		conv = &glm4MoeLiteModel{}
+	case "Lfm2ForCausalLM":
+		conv = &lfm2Model{}
 	default:
-		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
+		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}
 
 	if err := json.Unmarshal(bts, conv); err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	if t, ok := conv.(moreParser); ok {
 		if err := t.parseMore(fsys); err != nil {
-			return err
+			return nil, nil, err
 		}
 	}
 
 	t, err := parseTokenizer(fsys, conv.specialTokenTypes())
 	if err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
@@ -254,6 +353,19 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	default:
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}
+	return conv, t, nil
+}
+
+// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
+// and files it finds in the input path.
+// Supported input model formats include safetensors.
+// Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
+func ConvertModel(fsys fs.FS, f *os.File) error {
+	kv, t, err := LoadModelMetadata(fsys)
+	if err != nil {
+		return err
+	}
+	conv := kv.(ModelConverter)
 
 	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
 	if err != nil {
@@ -263,7 +375,7 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	return writeFile(f, conv.KV(t), conv.Tensors(ts))
 }
 
-func writeFile(f *os.File, kv ggml.KV, ts []*ggml.Tensor) error {
+func writeFile(f *os.File, kv KV, ts []*ggml.Tensor) error {
 	for i := range ts {
 		ts[i].Shape = slices.Clone(ts[i].Shape)
 		slices.Reverse(ts[i].Shape)

convert/convert_bert.go

@@ -88,7 +88,7 @@ func (p *bertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *bertModel) KV(t *Tokenizer) ggml.KV {
+func (p *bertModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "bert"
 	kv["bert.attention.causal"] = false

convert/convert_commandr.go

@@ -24,7 +24,7 @@ type commandrModel struct {
 
 var _ ModelConverter = (*commandrModel)(nil)
 
-func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
+func (p *commandrModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "command-r"
 	kv["general.name"] = "command-r"

convert/convert_deepseek2.go

@@ -47,7 +47,7 @@ type deepseek2Model struct {
 	Architecture string
 }
 
-func (p *deepseek2Model) KV(t *Tokenizer) ggml.KV {
+func (p *deepseek2Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "deepseek2"
 	kv["general.type"] = "model"

convert/convert_deepseekocr.go

@@ -41,7 +41,7 @@ type deepseekocr struct {
 	} `json:"vision_config"`
 }
 
-func (m *deepseekocr) KV(t *Tokenizer) ggml.KV {
+func (m *deepseekocr) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "deepseekocr"
 	kv["block_count"] = m.LanguageConfig.HiddenLayers

convert/convert_gemma.go

@@ -23,7 +23,7 @@ type gemmaModel struct {
 
 var _ ModelConverter = (*gemmaModel)(nil)
 
-func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
+func (p *gemmaModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma"
 	kv["gemma.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2.go

@@ -1,7 +1,5 @@
 package convert
 
-import "github.com/ollama/ollama/fs/ggml"
-
 type gemma2Model struct {
 	gemmaModel
 	SlidingWindow         uint32  `json:"sliding_window"`
@@ -9,7 +7,7 @@ type gemma2Model struct {
 	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
 }
 
-func (p *gemma2Model) KV(t *Tokenizer) ggml.KV {
+func (p *gemma2Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma2"
 	kv["gemma2.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2_adapter.go

@@ -6,6 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -15,7 +16,7 @@ type gemma2Adapter struct {
 
 var _ AdapterConverter = (*gemma2Adapter)(nil)
 
-func (p *gemma2Adapter) KV(baseKV ggml.KV) ggml.KV {
+func (p *gemma2Adapter) KV(baseKV fs.Config) KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "gemma2"
 	return kv

convert/convert_gemma3.go

@@ -3,8 +3,6 @@ package convert
 import (
 	"cmp"
 	"slices"
-
-	"github.com/ollama/ollama/fs/ggml"
 )
 
 type gemma3Model struct {
@@ -55,7 +53,7 @@ const (
 	gemma27BLayerCount = 62
 )
 
-func (p *gemma3Model) KV(t *Tokenizer) ggml.KV {
+func (p *gemma3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma3"
 

convert/convert_gemma3n.go

@@ -38,7 +38,7 @@ type gemma3nModel struct {
 	VisionModel struct{} `json:"vision_config"`
 }
 
-func (m *gemma3nModel) KV(t *Tokenizer) ggml.KV {
+func (m *gemma3nModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma3n"
 	kv["gemma3n.activation_sparsity_scale"] = slices.Collect(func(yield func(float32) bool) {

convert/convert_glm4moelite.go

@@ -0,0 +1,150 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"log/slog"
+	"regexp"
+	"strconv"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type glm4MoeLiteModel struct {
+	ModelParameters
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	HiddenLayers          uint32  `json:"num_hidden_layers"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+
+	RopeTheta     float32 `json:"rope_theta"`
+	QKNopeHeadDim uint32  `json:"qk_nope_head_dim"`
+	QKRopeHeadDim uint32  `json:"qk_rope_head_dim"`
+	KVLoraRank    uint32  `json:"kv_lora_rank"`
+	QLoraRank     uint32  `json:"q_lora_rank"`
+	VHeadDim      uint32  `json:"v_head_dim"`
+
+	ExpertCount            uint32  `json:"n_routed_experts"`
+	ExpertSharedCount      uint32  `json:"n_shared_experts"`
+	ExpertIntermediateSize uint32  `json:"moe_intermediate_size"`
+	ExpertUsedCount        uint32  `json:"num_experts_per_tok"`
+	ExpertWeightsNorm      bool    `json:"norm_topk_prob"`
+	ExpertWeightsScale     float32 `json:"routed_scaling_factor"`
+
+	LeadingDenseBlockCount uint32 `json:"first_k_dense_replace"`
+}
+
+func (p *glm4MoeLiteModel) KV(t *Tokenizer) KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "glm4moelite"
+	kv["general.type"] = "model"
+	kv["glm4moelite.block_count"] = p.HiddenLayers
+
+	numHeads := p.NumAttentionHeads
+	numKVHeads := p.NumKeyValueHeads
+
+	kv["glm4moelite.attention.head_count"] = numHeads
+	kv["glm4moelite.attention.head_count_kv"] = numKVHeads
+	kv["glm4moelite.attention.key_length"] = p.QKNopeHeadDim + p.QKRopeHeadDim
+	kv["glm4moelite.attention.kv_lora_rank"] = p.KVLoraRank
+	kv["glm4moelite.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["glm4moelite.attention.q_lora_rank"] = p.QLoraRank
+	kv["glm4moelite.attention.value_length"] = p.VHeadDim
+	kv["glm4moelite.context_length"] = p.MaxPositionEmbeddings
+	kv["glm4moelite.embedding_length"] = p.HiddenSize
+	kv["glm4moelite.expert_count"] = p.ExpertCount
+	kv["glm4moelite.expert_feed_forward_length"] = p.ExpertIntermediateSize
+	kv["glm4moelite.expert_shared_count"] = p.ExpertSharedCount
+
+	kv["glm4moelite.expert_gating_func"] = uint32(2)
+	kv["glm4moelite.expert_used_count"] = p.ExpertUsedCount
+	kv["glm4moelite.expert_weights_norm"] = p.ExpertWeightsNorm
+	kv["glm4moelite.expert_weights_scale"] = p.ExpertWeightsScale
+	kv["glm4moelite.feed_forward_length"] = p.IntermediateSize
+	kv["glm4moelite.leading_dense_block_count"] = p.LeadingDenseBlockCount
+
+	kv["glm4moelite.rope.dimension_count"] = p.QKRopeHeadDim
+	kv["glm4moelite.rope.freq_base"] = cmp.Or(p.RopeTheta, float32(1000000.0))
+
+	kv["tokenizer.ggml.pre"] = "glm4"
+
+	return kv
+}
+
+func (p *glm4MoeLiteModel) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.kv_a_proj_with_mqa", "attn_kv_a_mqa",
+		"self_attn.kv_a_layernorm", "attn_kv_a_norm",
+		"self_attn.kv_b_proj", "attn_kv_b",
+		"self_attn.q_a_proj", "attn_q_a",
+		"self_attn.q_a_layernorm", "attn_q_a_norm",
+		"self_attn.q_b_proj", "attn_q_b",
+		"self_attn.o_proj", "attn_output",
+		"post_attention_layernorm", "ffn_norm",
+		"mlp.shared_experts.down_proj", "ffn_down_shexp",
+		"mlp.shared_experts.gate_proj", "ffn_gate_shexp",
+		"mlp.shared_experts.up_proj", "ffn_up_shexp",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"mlp.gate.e_score_correction_bias", "exp_probs_b.bias",
+		"mlp.gate", "ffn_gate_inp",
+	}
+}
+
+func (p *glm4MoeLiteModel) Tensors(s []Tensor) (out []*ggml.Tensor) {
+	merges := make([]merge, p.HiddenLayers*3)
+	for i := range p.HiddenLayers {
+		merges[i*3+0] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.gate_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
+		}
+		merges[i*3+1] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.up_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
+		}
+		merges[i*3+2] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.down_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
+		}
+	}
+
+	skipLayer := func(n string, minValue uint32) bool {
+		re := regexp.MustCompile(`^blk\.(\d+)`)
+		matches := re.FindStringSubmatch(n)
+		if matches == nil {
+			return false
+		}
+
+		blkNum, err := strconv.Atoi(matches[1])
+		if err != nil {
+			return false
+		}
+
+		return uint32(blkNum) >= minValue
+	}
+
+	out, s = mergeTensors(s, merges...)
+	for _, t := range s {
+		// skip any additional layers (such as the Multi-Token Prediction layer)
+		if skipLayer(t.Name(), p.HiddenLayers) {
+			slog.Debug("skipping layer", "name", t.Name())
+			continue
+		}
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+	return out
+}

convert/convert_gptoss.go

@@ -37,7 +37,7 @@ type gptossModel struct {
 
 var _ ModelConverter = (*gptossModel)(nil)
 
-func (m *gptossModel) KV(t *Tokenizer) ggml.KV {
+func (m *gptossModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "gptoss"
 	kv["general.file_type"] = uint32(4)

convert/convert_lfm2.go

@@ -0,0 +1,100 @@
+package convert
+
+import (
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type lfm2Model struct {
+	ModelParameters
+	HiddenSize            uint32   `json:"hidden_size"`
+	NumHiddenLayers       uint32   `json:"num_hidden_layers"`
+	MaxPositionEmbeddings uint32   `json:"max_position_embeddings"`
+	IntermediateSize      uint32   `json:"intermediate_size"`
+	NumAttentionHeads     uint32   `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32   `json:"num_key_value_heads"`
+	RopeTheta             float32  `json:"rope_theta"`
+	NormEps               float32  `json:"norm_eps"`
+	ConvLCache            uint32   `json:"conv_L_cache"`
+	LayerTypes            []string `json:"layer_types"`
+	TieEmbedding          bool     `json:"tie_embedding"`
+}
+
+var _ ModelConverter = (*lfm2Model)(nil)
+
+func (p *lfm2Model) KV(t *Tokenizer) KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "lfm2"
+	kv["lfm2.vocab_size"] = p.VocabSize
+	kv["lfm2.block_count"] = p.NumHiddenLayers
+	kv["lfm2.embedding_length"] = p.HiddenSize
+	kv["lfm2.feed_forward_length"] = p.IntermediateSize
+	kv["lfm2.context_length"] = p.MaxPositionEmbeddings
+
+	// Build per-layer KV head count array based on layer_types
+	// (0 = shortconv layer, non-zero = attention layer with that many KV heads)
+	kvHeadCounts := make([]uint32, p.NumHiddenLayers)
+	for i := range p.NumHiddenLayers {
+		if int(i) < len(p.LayerTypes) && p.LayerTypes[i] == "full_attention" {
+			kvHeadCounts[i] = p.NumKeyValueHeads
+		}
+	}
+
+	kv["lfm2.attention.head_count"] = p.NumAttentionHeads
+	kv["lfm2.attention.head_count_kv"] = kvHeadCounts
+	kv["lfm2.attention.key_length"] = p.HiddenSize / p.NumAttentionHeads
+	kv["lfm2.attention.value_length"] = p.HiddenSize / p.NumAttentionHeads
+	kv["lfm2.attention.layer_norm_rms_epsilon"] = p.NormEps
+	kv["lfm2.rope.freq_base"] = p.RopeTheta
+	kv["lfm2.shortconv.l_cache"] = p.ConvLCache
+
+	return kv
+}
+
+func (p *lfm2Model) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+
+	for _, t := range ts {
+		shape := t.Shape()
+
+		// Squeeze conv weights: [D, 1, K] -> [D, K]
+		if strings.HasSuffix(t.Name(), "shortconv.conv.weight") {
+			if len(shape) == 3 && shape[1] == 1 {
+				shape = []uint64{shape[0], shape[2]}
+			}
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    slices.Clone(shape),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *lfm2Model) Replacements() []string {
+	return []string{
+		"model.embed_tokens", "token_embd",
+		"model.embedding_norm", "output_norm",
+		"model.layers", "blk",
+		"operator_norm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.out_proj", "attn_output",
+		"self_attn.q_layernorm", "attn_q_norm",
+		"self_attn.k_layernorm", "attn_k_norm",
+		"conv.conv", "shortconv.conv",
+		"conv.in_proj", "shortconv.in_proj",
+		"conv.out_proj", "shortconv.out_proj",
+		"feed_forward.w1", "ffn_gate",
+		"feed_forward.w2", "ffn_down",
+		"feed_forward.w3", "ffn_up",
+		"ffn_norm", "ffn_norm",
+	}
+}

convert/convert_llama.go

@@ -48,7 +48,7 @@ type llamaModel struct {
 
 var _ ModelConverter = (*llamaModel)(nil)
 
-func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
+func (p *llamaModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama"
 	kv["llama.vocab_size"] = p.VocabSize

convert/convert_llama4.go

@@ -35,7 +35,7 @@ type llama4Model struct {
 }
 
 // KV implements ModelConverter.
-func (p *llama4Model) KV(t *Tokenizer) ggml.KV {
+func (p *llama4Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama4"
 

convert/convert_llama_adapter.go

@@ -7,6 +7,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -18,13 +19,13 @@ type llamaAdapter struct {
 
 var _ AdapterConverter = (*llamaAdapter)(nil)
 
-func (p *llamaAdapter) KV(baseKV ggml.KV) ggml.KV {
+func (p *llamaAdapter) KV(baseKV fs.Config) KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "llama"
-	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
-	kv["llama.attention.head_count_kv"] = baseKV["llama.attention.head_count_kv"]
+	kv["llama.attention.head_count"] = baseKV.Value("llama.attention.head_count")
+	kv["llama.attention.head_count_kv"] = baseKV.Value("llama.attention.head_count_kv")
 
-	p.NumAttentionHeads = baseKV["llama.attention.head_count"].(uint32)
+	p.NumAttentionHeads = baseKV.Value("llama.attention.head_count").(uint32)
 
 	return kv
 }

convert/convert_mistral.go

@@ -60,7 +60,7 @@ type mistral3Model struct {
 	ProjectorHiddenAct      string `json:"projector_hidden_act"`
 }
 
-func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
+func (p *mistral3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "mistral3"
 	kv["mistral3.vocab_size"] = p.TextModel.VocabSize

convert/convert_mistral_causal.go

@@ -39,7 +39,7 @@ type mistral3CausalModel struct {
 	} `json:"rope_parameters"`
 }
 
-func (p *mistral3CausalModel) KV(t *Tokenizer) ggml.KV {
+func (p *mistral3CausalModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "mistral3"
 	kv["mistral3.vocab_size"] = p.VocabSize

convert/convert_mixtral.go

@@ -12,7 +12,7 @@ type mixtralModel struct {
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }
 
-func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
+func (p *mixtralModel) KV(t *Tokenizer) KV {
 	kv := p.llamaModel.KV(t)
 
 	if p.NumLocalExperts > 0 {

convert/convert_mllama.go

@@ -34,7 +34,7 @@ type mllamaModel struct {
 	} `json:"vision_config"`
 }
 
-func (m *mllamaModel) KV(t *Tokenizer) ggml.KV {
+func (m *mllamaModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "mllama"
 

convert/convert_nomicbert.go

@@ -87,7 +87,7 @@ func (p *nomicbertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *nomicbertModel) KV(t *Tokenizer) ggml.KV {
+func (p *nomicbertModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 
 	// Determine architecture based on MoE parameters (following qwen3 pattern)

convert/convert_olmo.go

@@ -34,7 +34,7 @@ type olmoModel struct {
 
 var _ ModelConverter = (*olmoModel)(nil)
 
-func (p *olmoModel) KV(t *Tokenizer) ggml.KV {
+func (p *olmoModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "olmo3"
 	kv["olmo3.block_count"] = p.NumHiddenLayers

convert/convert_phi3.go

@@ -37,7 +37,7 @@ type phi3Model struct {
 
 var _ ModelConverter = (*phi3Model)(nil)
 
-func (p *phi3Model) KV(t *Tokenizer) ggml.KV {
+func (p *phi3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "phi3"
 	kv["phi3.context_length"] = p.MaxPositionEmbeddings

convert/convert_qwen2.go

@@ -22,7 +22,7 @@ type qwen2Model struct {
 
 var _ ModelConverter = (*qwen2Model)(nil)
 
-func (q *qwen2Model) KV(t *Tokenizer) ggml.KV {
+func (q *qwen2Model) KV(t *Tokenizer) KV {
 	kv := q.ModelParameters.KV(t)
 	kv["general.architecture"] = "qwen2"
 	kv["qwen2.block_count"] = q.HiddenLayers

convert/convert_qwen25vl.go

@@ -29,7 +29,7 @@ type qwen25VLModel struct {
 
 var _ ModelConverter = (*qwen25VLModel)(nil)
 
-func (q *qwen25VLModel) KV(t *Tokenizer) ggml.KV {
+func (q *qwen25VLModel) KV(t *Tokenizer) KV {
 	kv := q.ModelParameters.KV(t)
 	kv["general.architecture"] = "qwen25vl"
 

convert/convert_qwen3.go

@@ -32,7 +32,7 @@ type qwen3Model struct {
 }
 
 // KV implements ModelConverter.
-func (q *qwen3Model) KV(t *Tokenizer) ggml.KV {
+func (q *qwen3Model) KV(t *Tokenizer) KV {
 	arch := "qwen3"
 	if q.NumExperts > 0 {
 		arch += "moe"

convert/convert_qwen3vl.go

@@ -45,7 +45,7 @@ func (m *qwen3VLModel) parseMore(fsys fs.FS) error {
 	return json.Unmarshal(bts, &m.VisionModel)
 }
 
-func (m *qwen3VLModel) KV(t *Tokenizer) ggml.KV {
+func (m *qwen3VLModel) KV(t *Tokenizer) KV {
 	kv := m.qwen3Model.KV(t)
 
 	arch := "qwen3vl"

convert/convert_test.go

@@ -19,6 +19,7 @@ import (
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
+	fsc "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -28,7 +29,7 @@ type tensorData struct {
 	Shape   []int  `json:"shape"`
 }
 
-func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, fsc.Config, ggml.Tensors) {
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), "f16")
@@ -59,9 +60,10 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	return r, m.KV(), m.Tensors()
 }
 
-func generateResultsJSON(t *testing.T, f *os.File, kv ggml.KV, tensors ggml.Tensors) map[string]string {
+func generateResultsJSON(t *testing.T, f *os.File, kv fsc.Config, tensors ggml.Tensors) map[string]string {
 	actual := make(map[string]string)
-	for k, v := range kv {
+	for k := range kv.Keys() {
+		v := kv.Value(k)
 		if s, ok := v.(json.Marshaler); !ok {
 			actual[k] = fmt.Sprintf("%v", v)
 		} else {
@@ -277,7 +279,7 @@ func generateSafetensorTestData(t *testing.T, tempDir string, tensorData map[str
 func TestConvertAdapter(t *testing.T) {
 	type AdapterCase struct {
 		Name     string
-		BaseKV   map[string]any
+		BaseKV   KV
 		Expected map[string]string
 	}
 

convert/reader.go

@@ -40,6 +40,7 @@ const (
 func (t tensorBase) Kind() uint32 {
 	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
 		strings.HasSuffix(t.name, ".bias") ||
+		strings.HasSuffix(t.name, ".shortconv.conv.weight") ||
 		t.name == "token_types.weight" ||
 		t.name == "v.positional_embedding_vlm" ||
 		t.name == "v.tile_position_embd.weight" ||

docs/README.md

@@ -14,6 +14,7 @@
 * [API Reference](https://docs.ollama.com/api)
 * [Modelfile Reference](https://docs.ollama.com/modelfile)
 * [OpenAI Compatibility](https://docs.ollama.com/api/openai-compatibility)
+* [Anthropic Compatibility](./api/anthropic-compatibility.mdx)
 
 ### Resources
 

docs/api.md

@@ -16,6 +16,7 @@
 - [Generate Embeddings](#generate-embeddings)
 - [List Running Models](#list-running-models)
 - [Version](#version)
+- [Experimental: Image Generation](#image-generation-experimental)
 
 ## Conventions
 
@@ -58,6 +59,15 @@ Advanced parameters (optional):
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 - `context` (deprecated): the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory
 
+Experimental image generation parameters (for image generation models only):
+
+> [!WARNING]
+> These parameters are experimental and may change in future versions.
+
+- `width`: width of the generated image in pixels
+- `height`: height of the generated image in pixels
+- `steps`: number of diffusion steps
+
 #### Structured outputs
 
 Structured outputs are supported by providing a JSON schema in the `format` parameter. The model will generate a response that matches the schema. See the [structured outputs](#request-structured-outputs) example below.
@@ -1867,3 +1877,55 @@ curl http://localhost:11434/api/version
   "version": "0.5.1"
 }
 ```
+
+## Experimental Features
+
+### Image Generation (Experimental)
+
+> [!WARNING]
+> Image generation is experimental and may change in future versions.
+
+Image generation is now supported through the standard `/api/generate` endpoint when using image generation models. The API automatically detects when an image generation model is being used.
+
+See the [Generate a completion](#generate-a-completion) section for the full API documentation. The experimental image generation parameters (`width`, `height`, `steps`) are documented there.
+
+#### Example
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/generate -d '{
+  "model": "x/z-image-turbo",
+  "prompt": "a sunset over mountains",
+  "width": 1024,
+  "height": 768
+}'
+```
+
+##### Response (streaming)
+
+Progress updates during generation:
+
+```json
+{
+  "model": "x/z-image-turbo",
+  "created_at": "2024-01-15T10:30:00.000000Z",
+  "completed": 5,
+  "total": 20,
+  "done": false
+}
+```
+
+##### Final Response
+
+```json
+{
+  "model": "x/z-image-turbo",
+  "created_at": "2024-01-15T10:30:15.000000Z",
+  "image": "iVBORw0KGgoAAAANSUhEUg...",
+  "done": true,
+  "done_reason": "stop",
+  "total_duration": 15000000000,
+  "load_duration": 2000000000
+}
+```

docs/api/anthropic-compatibility.mdx

@@ -0,0 +1,408 @@
+---
+title: Anthropic compatibility
+---
+
+Ollama provides compatibility with the [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) to help connect existing applications to Ollama, including tools like Claude Code.
+
+## Recommended models
+
+For coding use cases, models like `glm-4.7:cloud`, `minimax-m2.1:cloud`, and `qwen3-coder` are recommended.
+
+Pull a model before use:
+```shell
+ollama pull qwen3-coder
+ollama pull glm-4.7:cloud
+```
+
+## Usage
+
+### Environment variables
+
+To use Ollama with tools that expect the Anthropic API (like Claude Code), set these environment variables:
+
+```shell
+export ANTHROPIC_AUTH_TOKEN=ollama  # required but ignored
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama  # required but ignored
+```
+
+### Simple `/v1/messages` example
+
+<CodeGroup dropdown>
+
+```python basic.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',  # required but ignored
+)
+
+message = client.messages.create(
+    model='qwen3-coder',
+    max_tokens=1024,
+    messages=[
+        {'role': 'user', 'content': 'Hello, how are you?'}
+    ]
+)
+print(message.content[0].text)
+```
+
+```javascript basic.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama", // required but ignored
+});
+
+const message = await anthropic.messages.create({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  messages: [{ role: "user", content: "Hello, how are you?" }],
+});
+
+console.log(message.content[0].text);
+```
+
+```shell basic.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-H "x-api-key: ollama" \
+-H "anthropic-version: 2023-06-01" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "messages": [{ "role": "user", "content": "Hello, how are you?" }]
+}'
+```
+
+</CodeGroup>
+
+### Streaming example
+
+<CodeGroup dropdown>
+
+```python streaming.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',
+)
+
+with client.messages.stream(
+    model='qwen3-coder',
+    max_tokens=1024,
+    messages=[{'role': 'user', 'content': 'Count from 1 to 10'}]
+) as stream:
+    for text in stream.text_stream:
+        print(text, end='', flush=True)
+```
+
+```javascript streaming.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama",
+});
+
+const stream = await anthropic.messages.stream({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  messages: [{ role: "user", content: "Count from 1 to 10" }],
+});
+
+for await (const event of stream) {
+  if (
+    event.type === "content_block_delta" &&
+    event.delta.type === "text_delta"
+  ) {
+    process.stdout.write(event.delta.text);
+  }
+}
+```
+
+```shell streaming.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "stream": true,
+  "messages": [{ "role": "user", "content": "Count from 1 to 10" }]
+}'
+```
+
+</CodeGroup>
+
+### Tool calling example
+
+<CodeGroup dropdown>
+
+```python tools.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',
+)
+
+message = client.messages.create(
+    model='qwen3-coder',
+    max_tokens=1024,
+    tools=[
+        {
+            'name': 'get_weather',
+            'description': 'Get the current weather in a location',
+            'input_schema': {
+                'type': 'object',
+                'properties': {
+                    'location': {
+                        'type': 'string',
+                        'description': 'The city and state, e.g. San Francisco, CA'
+                    }
+                },
+                'required': ['location']
+            }
+        }
+    ],
+    messages=[{'role': 'user', 'content': "What's the weather in San Francisco?"}]
+)
+
+for block in message.content:
+    if block.type == 'tool_use':
+        print(f'Tool: {block.name}')
+        print(f'Input: {block.input}')
+```
+
+```javascript tools.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama",
+});
+
+const message = await anthropic.messages.create({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  tools: [
+    {
+      name: "get_weather",
+      description: "Get the current weather in a location",
+      input_schema: {
+        type: "object",
+        properties: {
+          location: {
+            type: "string",
+            description: "The city and state, e.g. San Francisco, CA",
+          },
+        },
+        required: ["location"],
+      },
+    },
+  ],
+  messages: [{ role: "user", content: "What's the weather in San Francisco?" }],
+});
+
+for (const block of message.content) {
+  if (block.type === "tool_use") {
+    console.log("Tool:", block.name);
+    console.log("Input:", block.input);
+  }
+}
+```
+
+```shell tools.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "tools": [
+    {
+      "name": "get_weather",
+      "description": "Get the current weather in a location",
+      "input_schema": {
+        "type": "object",
+        "properties": {
+          "location": {
+            "type": "string",
+            "description": "The city and state"
+          }
+        },
+        "required": ["location"]
+      }
+    }
+  ],
+  "messages": [{ "role": "user", "content": "What is the weather in San Francisco?" }]
+}'
+```
+
+</CodeGroup>
+
+## Using with Claude Code
+
+[Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
+
+```shell
+ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+```
+
+Or set the environment variables in your shell profile:
+
+```shell
+export ANTHROPIC_AUTH_TOKEN=ollama
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama
+```
+
+Then run Claude Code with any Ollama model:
+
+```shell
+# Local models
+claude --model qwen3-coder
+claude --model gpt-oss:20b
+
+# Cloud models
+claude --model glm-4.7:cloud
+claude --model minimax-m2.1:cloud
+```
+
+## Endpoints
+
+### `/v1/messages`
+
+#### Supported features
+
+- [x] Messages
+- [x] Streaming
+- [x] System prompts
+- [x] Multi-turn conversations
+- [x] Vision (images)
+- [x] Tools (function calling)
+- [x] Tool results
+- [x] Thinking/extended thinking
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `max_tokens`
+- [x] `messages`
+  - [x] Text `content`
+  - [x] Image `content` (base64)
+  - [x] Array of content blocks
+  - [x] `tool_use` blocks
+  - [x] `tool_result` blocks
+  - [x] `thinking` blocks
+- [x] `system` (string or array)
+- [x] `stream`
+- [x] `temperature`
+- [x] `top_p`
+- [x] `top_k`
+- [x] `stop_sequences`
+- [x] `tools`
+- [x] `thinking`
+- [ ] `tool_choice`
+- [ ] `metadata`
+
+#### Supported response fields
+
+- [x] `id`
+- [x] `type`
+- [x] `role`
+- [x] `model`
+- [x] `content` (text, tool_use, thinking blocks)
+- [x] `stop_reason` (end_turn, max_tokens, tool_use)
+- [x] `usage` (input_tokens, output_tokens)
+
+#### Streaming events
+
+- [x] `message_start`
+- [x] `content_block_start`
+- [x] `content_block_delta` (text_delta, input_json_delta, thinking_delta)
+- [x] `content_block_stop`
+- [x] `message_delta`
+- [x] `message_stop`
+- [x] `ping`
+- [x] `error`
+
+## Models
+
+Ollama supports both local and cloud models.
+
+### Local models
+
+Pull a local model before use:
+
+```shell
+ollama pull qwen3-coder
+```
+
+Recommended local models:
+- `qwen3-coder` - Excellent for coding tasks
+- `gpt-oss:20b` - Strong general-purpose model
+
+### Cloud models
+
+Cloud models are available immediately without pulling:
+
+- `glm-4.7:cloud` - High-performance cloud model
+- `minimax-m2.1:cloud` - Fast cloud model
+
+### Default model names
+
+For tooling that relies on default Anthropic model names such as `claude-3-5-sonnet`, use `ollama cp` to copy an existing model name:
+
+```shell
+ollama cp qwen3-coder claude-3-5-sonnet
+```
+
+Afterwards, this new model name can be specified in the `model` field:
+
+```shell
+curl http://localhost:11434/v1/messages \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "claude-3-5-sonnet",
+        "max_tokens": 1024,
+        "messages": [
+            {
+                "role": "user",
+                "content": "Hello!"
+            }
+        ]
+    }'
+```
+
+## Differences from the Anthropic API
+
+### Behavior differences
+
+- API key is accepted but not validated
+- `anthropic-version` header is accepted but not used
+- Token counts are approximations based on the underlying model's tokenizer
+
+### Not supported
+
+The following Anthropic API features are not currently supported:
+
+| Feature | Description |
+|---------|-------------|
+| `/v1/messages/count_tokens` | Token counting endpoint |
+| `tool_choice` | Forcing specific tool use or disabling tools |
+| `metadata` | Request metadata (user_id) |
+| Prompt caching | `cache_control` blocks for caching prefixes |
+| Batches API | `/v1/messages/batches` for async batch processing |
+| Citations | `citations` content blocks |
+| PDF support | `document` content blocks with PDF files |
+| Server-sent errors | `error` events during streaming (errors return HTTP status) |
+
+### Partial support
+
+| Feature | Status |
+|---------|--------|
+| Image content | Base64 images supported; URL images not supported |
+| Extended thinking | Basic support; `budget_tokens` accepted but not enforced |

docs/api/openai-compatibility.mdx

@@ -275,6 +275,73 @@ curl -X POST http://localhost:11434/v1/chat/completions \
 - [x] `dimensions`
 - [ ] `user`
 
+### `/v1/images/generations` (experimental)
+
+> Note: This endpoint is experimental and may change or be removed in future versions.
+
+Generate images using image generation models.
+
+<CodeGroup dropdown>
+
+```python images.py
+from openai import OpenAI
+
+client = OpenAI(
+    base_url='http://localhost:11434/v1/',
+    api_key='ollama',  # required but ignored
+)
+
+response = client.images.generate(
+    model='x/z-image-turbo',
+    prompt='A cute robot learning to paint',
+    size='1024x1024',
+    response_format='b64_json',
+)
+print(response.data[0].b64_json[:50] + '...')
+```
+
+```javascript images.js
+import OpenAI from "openai";
+
+const openai = new OpenAI({
+  baseURL: "http://localhost:11434/v1/",
+  apiKey: "ollama", // required but ignored
+});
+
+const response = await openai.images.generate({
+  model: "x/z-image-turbo",
+  prompt: "A cute robot learning to paint",
+  size: "1024x1024",
+  response_format: "b64_json",
+});
+
+console.log(response.data[0].b64_json.slice(0, 50) + "...");
+```
+
+```shell images.sh
+curl -X POST http://localhost:11434/v1/images/generations \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "x/z-image-turbo",
+  "prompt": "A cute robot learning to paint",
+  "size": "1024x1024",
+  "response_format": "b64_json"
+}'
+```
+
+</CodeGroup>
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `prompt`
+- [x] `size` (e.g. "1024x1024")
+- [x] `response_format` (only `b64_json` supported)
+- [ ] `n`
+- [ ] `quality`
+- [ ] `style`
+- [ ] `user`
+
 ### `/v1/responses`
 
 > Note: Added in Ollama v0.13.3

docs/capabilities/web-search.mdx

@@ -110,7 +110,7 @@ More Ollama [Python example](https://github.com/ollama/ollama-python/blob/main/e
 import { Ollama } from "ollama";
 
 const client = new Ollama();
-const results = await client.webSearch({ query: "what is ollama?" });
+const results = await client.webSearch("what is ollama?");
 console.log(JSON.stringify(results, null, 2));
 ```
 
@@ -213,7 +213,7 @@ models](https://ollama.com/models)\n\nAvailable for macOS, Windows, and Linux',
 import { Ollama } from "ollama";
 
 const client = new Ollama();
-const fetchResult = await client.webFetch({ url: "https://ollama.com" });
+const fetchResult = await client.webFetch("https://ollama.com");
 console.log(JSON.stringify(fetchResult, null, 2));
 ```
 

docs/docs.json

@@ -32,7 +32,9 @@
     "codeblocks": "system"
   },
   "contextual": {
-    "options": ["copy"]
+    "options": [
+      "copy"
+    ]
   },
   "navbar": {
     "links": [
@@ -52,7 +54,9 @@
       "display": "simple"
     },
     "examples": {
-      "languages": ["curl"]
+      "languages": [
+        "curl"
+      ]
     }
   },
   "redirects": [
@@ -97,6 +101,7 @@
           {
             "group": "Integrations",
             "pages": [
+              "/integrations/claude-code",
               "/integrations/vscode",
               "/integrations/jetbrains",
               "/integrations/codex",
@@ -106,7 +111,9 @@
               "/integrations/zed",
               "/integrations/roo-code",
               "/integrations/n8n",
-              "/integrations/xcode"
+              "/integrations/xcode",
+              "/integrations/onyx",
+              "/integrations/marimo"
             ]
           },
           {
@@ -139,7 +146,8 @@
               "/api/streaming",
               "/api/usage",
               "/api/errors",
-              "/api/openai-compatibility"
+              "/api/openai-compatibility",
+              "/api/anthropic-compatibility"
             ]
           },
           {

docs/faq.mdx

@@ -22,7 +22,7 @@ Please refer to the [GPU docs](./gpu).
 
 ## How can I specify the context window size?
 
-By default, Ollama uses a context window size of 2048 tokens.
+By default, Ollama uses a context window size of 4096 tokens.
 
 This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use:
 

docs/integrations/claude-code.mdx

@@ -0,0 +1,78 @@
+---
+title: Claude Code
+---
+
+Claude Code is Anthropic's agentic coding tool that can read, modify, and execute code in your working directory. 
+
+Open models can be used with Claude Code through Ollama's Anthropic-compatible API, enabling you to use models such as `qwen3-coder`, `gpt-oss:20b`, or other models.
+
+![Claude Code with Ollama](https://files.ollama.com/claude-code.png)
+
+## Install
+
+Install [Claude Code](https://code.claude.com/docs/en/overview):
+
+<CodeGroup>
+
+```shell macOS / Linux
+curl -fsSL https://claude.ai/install.sh | bash
+```
+
+```powershell Windows
+irm https://claude.ai/install.ps1 | iex
+```
+
+</CodeGroup>
+
+## Usage with Ollama
+
+Claude Code connects to Ollama using the Anthropic-compatible API.
+
+1. Set the environment variables:
+
+```shell
+export ANTHROPIC_AUTH_TOKEN=ollama
+export ANTHROPIC_BASE_URL=http://localhost:11434
+```
+
+2. Run Claude Code with an Ollama model:
+
+```shell
+claude --model gpt-oss:20b
+```
+
+Or run with environment variables inline:
+
+```shell
+ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 claude --model gpt-oss:20b
+```
+
+**Note:** Claude Code requires a large context window. We recommend at least 32K tokens. See the [context length documentation](/context-length) for how to adjust context length in Ollama.
+
+## Connecting to ollama.com
+
+1. Create an [API key](https://ollama.com/settings/keys) on ollama.com
+2. Set the environment variables:
+
+```shell
+export ANTHROPIC_BASE_URL=https://ollama.com
+export ANTHROPIC_API_KEY=<your-api-key>
+```
+
+3. Run Claude Code with a cloud model:
+
+```shell
+claude --model glm-4.7:cloud
+```
+
+## Recommended Models
+
+### Cloud models
+- `glm-4.7:cloud` - High-performance cloud model
+- `minimax-m2.1:cloud` - Fast cloud model
+- `qwen3-coder:480b` - Large coding model
+
+### Local models
+- `qwen3-coder` - Excellent for coding tasks
+- `gpt-oss:20b` - Strong general-purpose model
+- `gpt-oss:120b` - Larger general-purpose model for more complex tasks

docs/integrations/marimo.mdx

@@ -0,0 +1,73 @@
+---
+title: marimo
+---
+
+## Install
+
+Install [marimo](https://marimo.io). You can use `pip` or `uv` for this. You 
+can also use `uv` to create a sandboxed environment for marimo by running:
+
+```
+uvx marimo edit --sandbox notebook.py
+```
+
+## Usage with Ollama
+
+1. In marimo, go to the user settings and go to the AI tab. From here
+you can find and configure Ollama as an AI provider. For local use you
+would typically point the base url to `http://localhost:11434/v1`.
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-settings.png" 
+    alt="Ollama settings in marimo"
+    width="50%"
+  />
+</div>
+
+2. Once the AI provider is set up, you can turn on/off specific AI models you'd like to access. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-models.png" 
+    alt="Selecting an Ollama model"
+    width="50%"
+  />
+</div>
+
+3. You can also add a model to the list of available models by scrolling to the bottom and using the UI there. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-add-model.png" 
+    alt="Adding a new Ollama model"
+    width="50%"
+  />
+</div>
+
+4. Once configured, you can now use Ollama for AI chats in marimo.
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-chat.png" 
+    alt="Configure code completion"
+    width="50%"
+  />
+</div>
+
+4. Alternatively, you can now use Ollama for **inline code completion** in marimo. This can be configured in the "AI Features" tab. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-code-completion.png" 
+    alt="Configure code completion"
+    width="50%"
+  />
+</div>
+
+
+## Connecting to ollama.com
+
+1. Sign in to ollama cloud via `ollama signin` 
+2. In the ollama model settings add a model that ollama hosts, like `gpt-oss:120b`.
+3. You can now refer to this model in marimo!

docs/integrations/onyx.mdx

@@ -0,0 +1,63 @@
+---
+title: Onyx
+---
+
+## Overview
+[Onyx](http://onyx.app/) is a self-hostable Chat UI that integrates with all Ollama models. Features include:
+- Creating custom Agents
+- Web search
+- Deep Research
+- RAG over uploaded documents and connected apps
+- Connectors to applications like Google Drive, Email, Slack, etc.
+- MCP and OpenAPI Actions support
+- Image generation
+- User/Groups management, RBAC, SSO, etc.
+
+Onyx can be deployed for single users or large organizations.
+
+## Install Onyx
+
+Deploy Onyx with the [quickstart guide](https://docs.onyx.app/deployment/getting_started/quickstart).
+
+<Info>
+Resourcing/scaling docs [here](https://docs.onyx.app/deployment/getting_started/resourcing).
+</Info>
+
+## Usage with Ollama 
+
+1. Login to your Onyx deployment (create an account first).
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-login.png" 
+    alt="Onyx Login Page"
+    width="75%"
+  />
+</div>
+2. In the set-up process select `Ollama` as the LLM provider.
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-ollama-llm.png" 
+    alt="Onyx Set Up Form"
+    width="75%"
+  />
+</div>
+3. Provide your **Ollama API URL** and select your models.
+<Note>If you're running Onyx in Docker, to access your computer's local network use `http://host.docker.internal` instead of `http://127.0.0.1`.</Note>
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-ollama-form.png" 
+    alt="Selecting Ollama Models"
+    width="75%"
+  />
+</div>
+
+You can also easily connect up Onyx Cloud with the `Ollama Cloud` tab of the setup.
+
+## Send your first query
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-query.png" 
+    alt="Onyx Query Example"
+    width="75%"
+  />
+</div>

docs/linux.mdx

@@ -20,8 +20,8 @@ curl -fsSL https://ollama.com/install.sh | sh
 Download and extract the package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 Start Ollama:
@@ -41,8 +41,8 @@ ollama -v
 If you have an AMD GPU, also download and extract the additional ROCm package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 ### ARM64 install
@@ -50,8 +50,8 @@ curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz \
 Download and extract the ARM64-specific package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-arm64.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-arm64.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 ### Adding Ollama as a startup service (recommended)
@@ -146,8 +146,8 @@ curl -fsSL https://ollama.com/install.sh | sh
 Or by re-downloading Ollama:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 ## Installing specific versions

docs/troubleshooting.md

@@ -1,3 +0,0 @@
-# Troubleshooting
-
-For troubleshooting, see [https://docs.ollama.com/troubleshooting](https://docs.ollama.com/troubleshooting)

fs/config.go

@@ -1,5 +1,7 @@
 package fs
 
+import "iter"
+
 type Config interface {
 	Architecture() string
 	String(string, ...string) string
@@ -11,4 +13,8 @@ type Config interface {
 	Ints(string, ...[]int32) []int32
 	Floats(string, ...[]float32) []float32
 	Bools(string, ...[]bool) []bool
+
+	Len() int
+	Keys() iter.Seq[string]
+	Value(key string) any
 }

fs/ggml/ggml.go

@@ -6,7 +6,9 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"iter"
 	"log/slog"
+	"maps"
 	"math"
 	"slices"
 	"strings"
@@ -239,6 +241,18 @@ func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
 	return val.values
 }
 
+func (kv KV) Len() int {
+	return len(kv)
+}
+
+func (kv KV) Keys() iter.Seq[string] {
+	return maps.Keys(kv)
+}
+
+func (kv KV) Value(key string) any {
+	return kv[key]
+}
+
 func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
 		"bert",
@@ -255,6 +269,8 @@ func (kv KV) OllamaEngineRequired() bool {
 		"qwen25vl",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
+		"glm4moelite",
+		"lfm2",
 	}, kv.Architecture())
 }
 
@@ -842,7 +858,9 @@ func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
 		"bert",
 		"gemma3",
+		"glm4moelite",
 		"gptoss", "gpt-oss",
+		"lfm2",
 		"mistral3",
 		"olmo3",
 		"qwen3", "qwen3moe",

fs/ggml/gguf.go

@@ -8,12 +8,12 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
-	"maps"
 	"os"
 	"runtime"
 	"slices"
 	"strings"
 
+	"github.com/ollama/ollama/fs"
 	"golang.org/x/sync/errgroup"
 )
 
@@ -508,7 +508,7 @@ func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
 	return binary.Write(w, binary.LittleEndian, s)
 }
 
-func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
+func WriteGGUF(f *os.File, kv fs.Config, ts []*Tensor) error {
 	arch := kv.String("general.architecture")
 	if arch == "" {
 		return fmt.Errorf("architecture not set")
@@ -526,12 +526,12 @@ func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 		return err
 	}
 
-	if err := binary.Write(f, binary.LittleEndian, uint64(len(kv))); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, uint64(kv.Len())); err != nil {
 		return err
 	}
 
-	for _, key := range slices.Sorted(maps.Keys(kv)) {
-		if err := ggufWriteKV(f, arch, key, kv[key]); err != nil {
+	for _, key := range slices.Sorted(kv.Keys()) {
+		if err := ggufWriteKV(f, arch, key, kv.Value(key)); err != nil {
 			return err
 		}
 	}

integration/imagegen_test.go

@@ -0,0 +1,148 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"encoding/base64"
+	"fmt"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestImageGeneration(t *testing.T) {
+	skipUnderMinVRAM(t, 8)
+
+	type testCase struct {
+		imageGenModel string
+		visionModel   string
+		prompt        string
+		expectedWords []string
+	}
+
+	testCases := []testCase{
+		{
+			imageGenModel: "jmorgan/z-image-turbo",
+			visionModel:   "llama3.2-vision",
+			prompt:        "A cartoon style llama flying like a superhero through the air with clouds in the background",
+			expectedWords: []string{"llama", "flying", "cartoon", "cloud", "sky", "superhero", "air", "animal", "camelid"},
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(fmt.Sprintf("%s->%s", tc.imageGenModel, tc.visionModel), func(t *testing.T) {
+			ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute)
+			defer cancel()
+
+			client, _, cleanup := InitServerConnection(ctx, t)
+			defer cleanup()
+
+			// Pull both models
+			if err := PullIfMissing(ctx, client, tc.imageGenModel); err != nil {
+				t.Fatalf("failed to pull image gen model: %v", err)
+			}
+			if err := PullIfMissing(ctx, client, tc.visionModel); err != nil {
+				t.Fatalf("failed to pull vision model: %v", err)
+			}
+
+			// Generate the image
+			t.Logf("Generating image with prompt: %s", tc.prompt)
+			imageBase64, err := generateImage(ctx, client, tc.imageGenModel, tc.prompt)
+			if err != nil {
+				if strings.Contains(err.Error(), "image generation not available") {
+					t.Skip("Target system does not support image generation")
+				} else if strings.Contains(err.Error(), "executable file not found in") { // Windows pattern, not yet supported
+					t.Skip("Windows does not support image generation yet")
+				} else if strings.Contains(err.Error(), "CUDA driver version is insufficient") {
+					t.Skip("Driver is too old")
+				} else if strings.Contains(err.Error(), "insufficient memory for image generation") {
+					t.Skip("insufficient memory for image generation")
+				} else if strings.Contains(err.Error(), "error while loading shared libraries: libcuda.so.1") { // AMD GPU or CPU
+					t.Skip("CUDA GPU is not available")
+				} else if strings.Contains(err.Error(), "ollama-mlx: no such file or directory") {
+					// most likely linux arm - not supported yet
+					t.Skip("unsupported architecture")
+				}
+				t.Fatalf("failed to generate image: %v", err)
+			}
+
+			imageData, err := base64.StdEncoding.DecodeString(imageBase64)
+			if err != nil {
+				t.Fatalf("failed to decode image: %v", err)
+			}
+			t.Logf("Generated image: %d bytes", len(imageData))
+
+			// Preload vision model and check GPU loading
+			err = client.Generate(ctx, &api.GenerateRequest{Model: tc.visionModel}, func(response api.GenerateResponse) error { return nil })
+			if err != nil {
+				t.Fatalf("failed to load vision model: %v", err)
+			}
+
+			// Use vision model to describe the image
+			chatReq := api.ChatRequest{
+				Model: tc.visionModel,
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Describe this image in detail. What is shown? What style is it? What is the main subject doing?",
+						Images:  []api.ImageData{imageData},
+					},
+				},
+				Stream: &stream,
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			}
+
+			// Verify the vision model's response contains expected keywords
+			response := DoChat(ctx, t, client, chatReq, tc.expectedWords, 240*time.Second, 30*time.Second)
+			if response != nil {
+				t.Logf("Vision model response: %s", response.Content)
+
+				// Additional detailed check for keywords
+				content := strings.ToLower(response.Content)
+				foundWords := []string{}
+				missingWords := []string{}
+				for _, word := range tc.expectedWords {
+					if strings.Contains(content, word) {
+						foundWords = append(foundWords, word)
+					} else {
+						missingWords = append(missingWords, word)
+					}
+				}
+				t.Logf("Found keywords: %v", foundWords)
+				if len(missingWords) > 0 {
+					t.Logf("Missing keywords (at least one was found so test passed): %v", missingWords)
+				}
+			}
+		})
+	}
+}
+
+// generateImage calls the Ollama API to generate an image and returns the base64 image data
+func generateImage(ctx context.Context, client *api.Client, model, prompt string) (string, error) {
+	var imageBase64 string
+
+	err := client.Generate(ctx, &api.GenerateRequest{
+		Model:  model,
+		Prompt: prompt,
+	}, func(resp api.GenerateResponse) error {
+		if resp.Image != "" {
+			imageBase64 = resp.Image
+		}
+		return nil
+	})
+	if err != nil {
+		return "", fmt.Errorf("failed to generate image: %w", err)
+	}
+
+	if imageBase64 == "" {
+		return "", fmt.Errorf("no image data in response")
+	}
+
+	return imageBase64, nil
+}

integration/tools_test.go

@@ -131,7 +131,7 @@ func TestAPIToolCalling(t *testing.T) {
 					t.Errorf("unexpected tool called: got %q want %q", lastToolCall.Function.Name, "get_weather")
 				}
 
-				if _, ok := lastToolCall.Function.Arguments["location"]; !ok {
+				if _, ok := lastToolCall.Function.Arguments.Get("location"); !ok {
 					t.Errorf("expected tool arguments to include 'location', got: %s", lastToolCall.Function.Arguments.String())
 				}
 			case <-ctx.Done():

integration/utils_test.go

@@ -38,6 +38,7 @@ var (
 
 	// Note: add newer models at the top of the list to test them first
 	ollamaEngineChatModels = []string{
+		"lfm2.5-thinking",
 		"ministral-3",
 		"qwen3-coder:30b",
 		"gpt-oss:20b",
@@ -143,6 +144,7 @@ var (
 		"granite3.3",
 		"hermes3",
 		"internlm2",
+		"lfm2.5-thinking",
 		"llama-guard3",
 		"llama-pro",
 		"llama2-chinese",
@@ -263,6 +265,7 @@ var (
 		"snowflake-arctic-embed2",
 	}
 	libraryToolsModels = []string{
+		"lfm2.5-thinking",
 		"qwen3-vl",
 		"gpt-oss:20b",
 		"gpt-oss:120b",

llm/server.go

@@ -1464,6 +1464,12 @@ type CompletionRequest struct {
 
 	// TopLogprobs specifies the number of most likely alternative tokens to return (0-20)
 	TopLogprobs int
+
+	// Image generation fields
+	Width  int32 `json:"width,omitempty"`
+	Height int32 `json:"height,omitempty"`
+	Steps  int32 `json:"steps,omitempty"`
+	Seed   int64 `json:"seed,omitempty"`
 }
 
 // DoneReason represents the reason why a completion response is done
@@ -1512,6 +1518,15 @@ type CompletionResponse struct {
 
 	// Logprobs contains log probability information if requested
 	Logprobs []Logprob `json:"logprobs,omitempty"`
+
+	// Image contains base64-encoded image data for image generation
+	Image string `json:"image,omitempty"`
+
+	// Step is the current step in image generation
+	Step int `json:"step,omitempty"`
+
+	// TotalSteps is the total number of steps for image generation
+	TotalSteps int `json:"total_steps,omitempty"`
 }
 
 func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error {

manifest/layer.go

@@ -1,4 +1,4 @@
-package server
+package manifest
 
 import (
 	"crypto/sha256"
@@ -13,11 +13,16 @@ type Layer struct {
 	Digest    string `json:"digest"`
 	Size      int64  `json:"size"`
 	From      string `json:"from,omitempty"`
-	status    string
+	Name      string `json:"name,omitempty"` // tensor name, e.g., "text_encoder/model.embed_tokens.weight"
+	Status    string `json:"-"`
 }
 
+const (
+	MediaTypeImageTensor = "application/vnd.ollama.image.tensor"
+)
+
 func NewLayer(r io.Reader, mediatype string) (Layer, error) {
-	blobs, err := GetBlobsPath("")
+	blobs, err := BlobsPath("")
 	if err != nil {
 		return Layer{}, err
 	}
@@ -40,7 +45,7 @@ func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 	}
 
 	digest := fmt.Sprintf("sha256:%x", sha256sum.Sum(nil))
-	blob, err := GetBlobsPath(digest)
+	blob, err := BlobsPath(digest)
 	if err != nil {
 		return Layer{}, err
 	}
@@ -60,7 +65,7 @@ func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 		MediaType: mediatype,
 		Digest:    digest,
 		Size:      n,
-		status:    fmt.Sprintf("%s %s", status, digest),
+		Status:    fmt.Sprintf("%s %s", status, digest),
 	}, nil
 }
 
@@ -69,7 +74,7 @@ func NewLayerFromLayer(digest, mediatype, from string) (Layer, error) {
 		return Layer{}, errors.New("creating new layer from layer with empty digest")
 	}
 
-	blob, err := GetBlobsPath(digest)
+	blob, err := BlobsPath(digest)
 	if err != nil {
 		return Layer{}, err
 	}
@@ -84,7 +89,7 @@ func NewLayerFromLayer(digest, mediatype, from string) (Layer, error) {
 		Digest:    digest,
 		Size:      fi.Size(),
 		From:      from,
-		status:    fmt.Sprintf("using existing layer %s", digest),
+		Status:    fmt.Sprintf("using existing layer %s", digest),
 	}, nil
 }
 
@@ -93,7 +98,7 @@ func (l *Layer) Open() (io.ReadSeekCloser, error) {
 		return nil, errors.New("opening layer with empty digest")
 	}
 
-	blob, err := GetBlobsPath(l.Digest)
+	blob, err := BlobsPath(l.Digest)
 	if err != nil {
 		return nil, err
 	}
@@ -121,7 +126,7 @@ func (l *Layer) Remove() error {
 		}
 	}
 
-	blob, err := GetBlobsPath(l.Digest)
+	blob, err := BlobsPath(l.Digest)
 	if err != nil {
 		return err
 	}

manifest/manifest.go

@@ -1,10 +1,9 @@
-package server
+package manifest
 
 import (
 	"crypto/sha256"
 	"encoding/hex"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -33,12 +32,38 @@ func (m *Manifest) Size() (size int64) {
 	return
 }
 
+func (m *Manifest) Digest() string {
+	return m.digest
+}
+
+func (m *Manifest) FileInfo() os.FileInfo {
+	return m.fi
+}
+
+// ReadConfigJSON reads and unmarshals a config layer as JSON.
+func (m *Manifest) ReadConfigJSON(configPath string, v any) error {
+	for _, layer := range m.Layers {
+		if layer.MediaType == "application/vnd.ollama.image.json" && layer.Name == configPath {
+			blobPath, err := BlobsPath(layer.Digest)
+			if err != nil {
+				return err
+			}
+			data, err := os.ReadFile(blobPath)
+			if err != nil {
+				return err
+			}
+			return json.Unmarshal(data, v)
+		}
+	}
+	return fmt.Errorf("config %q not found in manifest", configPath)
+}
+
 func (m *Manifest) Remove() error {
 	if err := os.Remove(m.filepath); err != nil {
 		return err
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return err
 	}
@@ -47,16 +72,40 @@ func (m *Manifest) Remove() error {
 }
 
 func (m *Manifest) RemoveLayers() error {
-	for _, layer := range append(m.Layers, m.Config) {
-		if layer.Digest != "" {
-			if err := layer.Remove(); errors.Is(err, os.ErrNotExist) {
-				slog.Debug("layer does not exist", "digest", layer.Digest)
-			} else if err != nil {
-				return err
+	ms, err := Manifests(true)
+	if err != nil {
+		return err
+	}
+
+	// Build set of digests still in use by other manifests
+	inUse := make(map[string]struct{})
+	for _, other := range ms {
+		for _, layer := range append(other.Layers, other.Config) {
+			if layer.Digest != "" {
+				inUse[layer.Digest] = struct{}{}
 			}
 		}
 	}
 
+	// Remove layers not used by any other manifest
+	for _, layer := range append(m.Layers, m.Config) {
+		if layer.Digest == "" {
+			continue
+		}
+		if _, used := inUse[layer.Digest]; used {
+			continue
+		}
+		blob, err := BlobsPath(layer.Digest)
+		if err != nil {
+			return err
+		}
+		if err := os.Remove(blob); os.IsNotExist(err) {
+			slog.Debug("layer does not exist", "digest", layer.Digest)
+		} else if err != nil {
+			return err
+		}
+	}
+
 	return nil
 }
 
@@ -65,7 +114,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 		return nil, model.Unqualified(n)
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return nil, err
 	}
@@ -97,7 +146,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 }
 
 func WriteManifest(name model.Name, config Layer, layers []Layer) error {
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return err
 	}
@@ -124,7 +173,7 @@ func WriteManifest(name model.Name, config Layer, layers []Layer) error {
 }
 
 func Manifests(continueOnError bool) (map[model.Name]*Manifest, error) {
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return nil, err
 	}

manifest/manifest_test.go

@@ -1,4 +1,4 @@
-package server
+package manifest
 
 import (
 	"encoding/json"

manifest/paths.go

@@ -0,0 +1,95 @@
+package manifest
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/types/model"
+)
+
+var ErrInvalidDigestFormat = errors.New("invalid digest format")
+
+func Path() (string, error) {
+	path := filepath.Join(envconfig.Models(), "manifests")
+	if err := os.MkdirAll(path, 0o755); err != nil {
+		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
+	}
+
+	return path, nil
+}
+
+// PathForName returns the path to the manifest file for a specific model name.
+func PathForName(n model.Name) (string, error) {
+	if !n.IsValid() {
+		return "", os.ErrNotExist
+	}
+
+	manifests, err := Path()
+	if err != nil {
+		return "", err
+	}
+
+	return filepath.Join(manifests, n.Filepath()), nil
+}
+
+func BlobsPath(digest string) (string, error) {
+	// only accept actual sha256 digests
+	pattern := "^sha256[:-][0-9a-fA-F]{64}$"
+	re := regexp.MustCompile(pattern)
+
+	if digest != "" && !re.MatchString(digest) {
+		return "", ErrInvalidDigestFormat
+	}
+
+	digest = strings.ReplaceAll(digest, ":", "-")
+	path := filepath.Join(envconfig.Models(), "blobs", digest)
+	dirPath := filepath.Dir(path)
+	if digest == "" {
+		dirPath = path
+	}
+
+	if err := os.MkdirAll(dirPath, 0o755); err != nil {
+		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
+	}
+
+	return path, nil
+}
+
+// PruneDirectory removes empty directories recursively.
+func PruneDirectory(path string) error {
+	info, err := os.Lstat(path)
+	if err != nil {
+		return err
+	}
+
+	if info.IsDir() && info.Mode()&os.ModeSymlink == 0 {
+		entries, err := os.ReadDir(path)
+		if err != nil {
+			return err
+		}
+
+		for _, entry := range entries {
+			if err := PruneDirectory(filepath.Join(path, entry.Name())); err != nil {
+				return err
+			}
+		}
+
+		entries, err = os.ReadDir(path)
+		if err != nil {
+			return err
+		}
+
+		if len(entries) > 0 {
+			return nil
+		}
+
+		return os.Remove(path)
+	}
+
+	return nil
+}

middleware/anthropic.go

@@ -0,0 +1,152 @@
+package middleware
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/ollama/ollama/anthropic"
+	"github.com/ollama/ollama/api"
+)
+
+// AnthropicWriter wraps the response writer to transform Ollama responses to Anthropic format
+type AnthropicWriter struct {
+	BaseWriter
+	stream    bool
+	id        string
+	model     string
+	converter *anthropic.StreamConverter
+}
+
+func (w *AnthropicWriter) writeError(data []byte) (int, error) {
+	var errData struct {
+		Error string `json:"error"`
+	}
+	if err := json.Unmarshal(data, &errData); err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err := json.NewEncoder(w.ResponseWriter).Encode(anthropic.NewError(w.ResponseWriter.Status(), errData.Error))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *AnthropicWriter) writeEvent(eventType string, data any) error {
+	d, err := json.Marshal(data)
+	if err != nil {
+		return err
+	}
+	_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("event: %s\ndata: %s\n\n", eventType, d)))
+	if err != nil {
+		return err
+	}
+	if f, ok := w.ResponseWriter.(http.Flusher); ok {
+		f.Flush()
+	}
+	return nil
+}
+
+func (w *AnthropicWriter) writeResponse(data []byte) (int, error) {
+	var chatResponse api.ChatResponse
+	err := json.Unmarshal(data, &chatResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	if w.stream {
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+
+		events := w.converter.Process(chatResponse)
+		for _, event := range events {
+			if err := w.writeEvent(event.Event, event.Data); err != nil {
+				return 0, err
+			}
+		}
+		return len(data), nil
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	response := anthropic.ToMessagesResponse(w.id, chatResponse)
+	return len(data), json.NewEncoder(w.ResponseWriter).Encode(response)
+}
+
+func (w *AnthropicWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+// AnthropicMessagesMiddleware handles Anthropic Messages API requests
+func AnthropicMessagesMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req anthropic.MessagesRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Model == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "model is required"))
+			return
+		}
+
+		if req.MaxTokens <= 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "max_tokens is required and must be positive"))
+			return
+		}
+
+		if len(req.Messages) == 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "messages is required"))
+			return
+		}
+
+		chatReq, err := anthropic.FromMessagesRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		// Set think to nil when being used with Anthropic API to connect to tools like claude code
+		c.Set("relax_thinking", true)
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, anthropic.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		messageID := anthropic.GenerateMessageID()
+
+		w := &AnthropicWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			stream:     req.Stream,
+			id:         messageID,
+			model:      req.Model,
+			converter:  anthropic.NewStreamConverter(messageID, req.Model),
+		}
+
+		if req.Stream {
+			c.Writer.Header().Set("Content-Type", "text/event-stream")
+			c.Writer.Header().Set("Cache-Control", "no-cache")
+			c.Writer.Header().Set("Connection", "keep-alive")
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}

middleware/anthropic_test.go

@@ -0,0 +1,607 @@
+package middleware
+
+import (
+	"bytes"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/go-cmp/cmp"
+	"github.com/google/go-cmp/cmp/cmpopts"
+
+	"github.com/ollama/ollama/anthropic"
+	"github.com/ollama/ollama/api"
+)
+
+func captureAnthropicRequest(capturedRequest any) gin.HandlerFunc {
+	return func(c *gin.Context) {
+		bodyBytes, _ := io.ReadAll(c.Request.Body)
+		c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+		_ = json.Unmarshal(bodyBytes, capturedRequest)
+		c.Next()
+	}
+}
+
+// testProps creates ToolPropertiesMap from a map (convenience function for tests)
+func testProps(m map[string]api.ToolProperty) *api.ToolPropertiesMap {
+	props := api.NewToolPropertiesMap()
+	for k, v := range m {
+		props.Set(k, v)
+	}
+	return props
+}
+
+func TestAnthropicMessagesMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.ChatRequest
+		err  anthropic.ErrorResponse
+	}
+
+	var capturedRequest *api.ChatRequest
+	stream := true
+
+	testCases := []testCase{
+		{
+			name: "basic message",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with system prompt",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"system": "You are helpful.",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "system", Content: "You are helpful."},
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with options",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 2048,
+				"temperature": 0.7,
+				"top_p": 0.9,
+				"top_k": 40,
+				"stop_sequences": ["\n", "END"],
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{
+					"num_predict": 2048,
+					"temperature": 0.7,
+					"top_p":       0.9,
+					"top_k":       40,
+					"stop":        []string{"\n", "END"},
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "streaming",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"stream": true,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &stream,
+			},
+		},
+		{
+			name: "with tools",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "What's the weather?"}
+				],
+				"tools": [{
+					"name": "get_weather",
+					"description": "Get current weather",
+					"input_schema": {
+						"type": "object",
+						"properties": {
+							"location": {"type": "string"}
+						},
+						"required": ["location"]
+					}
+				}]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "What's the weather?"},
+				},
+				Tools: []api.Tool{
+					{
+						Type: "function",
+						Function: api.ToolFunction{
+							Name:        "get_weather",
+							Description: "Get current weather",
+							Parameters: api.ToolFunctionParameters{
+								Type:     "object",
+								Required: []string{"location"},
+								Properties: testProps(map[string]api.ToolProperty{
+									"location": {Type: api.PropertyType{"string"}},
+								}),
+							},
+						},
+					},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with tool result",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "What's the weather?"},
+					{"role": "assistant", "content": [
+						{"type": "tool_use", "id": "call_123", "name": "get_weather", "input": {"location": "Paris"}}
+					]},
+					{"role": "user", "content": [
+						{"type": "tool_result", "tool_use_id": "call_123", "content": "Sunny, 22°C"}
+					]}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "What's the weather?"},
+					{
+						Role: "assistant",
+						ToolCalls: []api.ToolCall{
+							{
+								ID: "call_123",
+								Function: api.ToolCallFunction{
+									Name:      "get_weather",
+									Arguments: testArgs(map[string]any{"location": "Paris"}),
+								},
+							},
+						},
+					},
+					{Role: "tool", Content: "Sunny, 22°C", ToolCallID: "call_123"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with thinking enabled",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"thinking": {"type": "enabled", "budget_tokens": 1000},
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+				Think:   &api.ThinkValue{Value: true},
+			},
+		},
+		{
+			name: "missing model error",
+			body: `{
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "model is required",
+				},
+			},
+		},
+		{
+			name: "missing max_tokens error",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "max_tokens is required and must be positive",
+				},
+			},
+		},
+		{
+			name: "missing messages error",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "messages is required",
+				},
+			},
+		},
+		{
+			name: "tool_use missing id error",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "assistant", "content": [
+						{"type": "tool_use", "name": "test"}
+					]}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "tool_use block missing required 'id' field",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware(), captureAnthropicRequest(&capturedRequest))
+	router.Handle(http.MethodPost, "/v1/messages", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if tc.err.Type != "" {
+				// Expect error
+				if resp.Code == http.StatusOK {
+					t.Fatalf("expected error response, got 200 OK")
+				}
+				var errResp anthropic.ErrorResponse
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatalf("failed to unmarshal error: %v", err)
+				}
+				if errResp.Type != tc.err.Type {
+					t.Errorf("expected error type %q, got %q", tc.err.Type, errResp.Type)
+				}
+				if errResp.Error.Type != tc.err.Error.Type {
+					t.Errorf("expected error.type %q, got %q", tc.err.Error.Type, errResp.Error.Type)
+				}
+				if errResp.Error.Message != tc.err.Error.Message {
+					t.Errorf("expected error.message %q, got %q", tc.err.Error.Message, errResp.Error.Message)
+				}
+				return
+			}
+
+			if resp.Code != http.StatusOK {
+				t.Fatalf("unexpected status code: %d, body: %s", resp.Code, resp.Body.String())
+			}
+
+			if capturedRequest == nil {
+				t.Fatal("request was not captured")
+			}
+
+			// Compare relevant fields
+			if capturedRequest.Model != tc.req.Model {
+				t.Errorf("model mismatch: got %q, want %q", capturedRequest.Model, tc.req.Model)
+			}
+
+			if diff := cmp.Diff(tc.req.Messages, capturedRequest.Messages,
+				cmpopts.IgnoreUnexported(api.ToolCallFunctionArguments{}, api.ToolPropertiesMap{})); diff != "" {
+				t.Errorf("messages mismatch (-want +got):\n%s", diff)
+			}
+
+			if tc.req.Stream != nil && capturedRequest.Stream != nil {
+				if *tc.req.Stream != *capturedRequest.Stream {
+					t.Errorf("stream mismatch: got %v, want %v", *capturedRequest.Stream, *tc.req.Stream)
+				}
+			}
+
+			if tc.req.Think != nil {
+				if capturedRequest.Think == nil {
+					t.Error("expected Think to be set")
+				} else if capturedRequest.Think.Value != tc.req.Think.Value {
+					t.Errorf("Think mismatch: got %v, want %v", capturedRequest.Think.Value, tc.req.Think.Value)
+				}
+			}
+		})
+	}
+}
+
+func TestAnthropicMessagesMiddleware_Headers(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	t.Run("streaming sets correct headers", func(t *testing.T) {
+		router := gin.New()
+		router.Use(AnthropicMessagesMiddleware())
+		router.POST("/v1/messages", func(c *gin.Context) {
+			// Check headers were set
+			if c.Writer.Header().Get("Content-Type") != "text/event-stream" {
+				t.Errorf("expected Content-Type text/event-stream, got %q", c.Writer.Header().Get("Content-Type"))
+			}
+			if c.Writer.Header().Get("Cache-Control") != "no-cache" {
+				t.Errorf("expected Cache-Control no-cache, got %q", c.Writer.Header().Get("Cache-Control"))
+			}
+			c.Status(http.StatusOK)
+		})
+
+		body := `{"model": "test", "max_tokens": 100, "stream": true, "messages": [{"role": "user", "content": "Hi"}]}`
+		req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+		req.Header.Set("Content-Type", "application/json")
+
+		resp := httptest.NewRecorder()
+		router.ServeHTTP(resp, req)
+	})
+}
+
+func TestAnthropicMessagesMiddleware_InvalidJSON(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware())
+	router.POST("/v1/messages", func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	})
+
+	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(`{invalid json`))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusBadRequest {
+		t.Errorf("expected status 400, got %d", resp.Code)
+	}
+
+	var errResp anthropic.ErrorResponse
+	if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+		t.Fatalf("failed to unmarshal error: %v", err)
+	}
+
+	if errResp.Type != "error" {
+		t.Errorf("expected type 'error', got %q", errResp.Type)
+	}
+	if errResp.Error.Type != "invalid_request_error" {
+		t.Errorf("expected error type 'invalid_request_error', got %q", errResp.Error.Type)
+	}
+}
+
+func TestAnthropicWriter_NonStreaming(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware())
+	router.POST("/v1/messages", func(c *gin.Context) {
+		// Simulate Ollama response
+		resp := api.ChatResponse{
+			Model: "test-model",
+			Message: api.Message{
+				Role:    "assistant",
+				Content: "Hello there!",
+			},
+			Done:       true,
+			DoneReason: "stop",
+			Metrics: api.Metrics{
+				PromptEvalCount: 10,
+				EvalCount:       5,
+			},
+		}
+		data, _ := json.Marshal(resp)
+		c.Writer.WriteHeader(http.StatusOK)
+		_, _ = c.Writer.Write(data)
+	})
+
+	body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
+	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", resp.Code)
+	}
+
+	var result anthropic.MessagesResponse
+	if err := json.Unmarshal(resp.Body.Bytes(), &result); err != nil {
+		t.Fatalf("failed to unmarshal response: %v", err)
+	}
+
+	if result.Type != "message" {
+		t.Errorf("expected type 'message', got %q", result.Type)
+	}
+	if result.Role != "assistant" {
+		t.Errorf("expected role 'assistant', got %q", result.Role)
+	}
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Text == nil || *result.Content[0].Text != "Hello there!" {
+		t.Errorf("expected text 'Hello there!', got %v", result.Content[0].Text)
+	}
+	if result.StopReason != "end_turn" {
+		t.Errorf("expected stop_reason 'end_turn', got %q", result.StopReason)
+	}
+	if result.Usage.InputTokens != 10 {
+		t.Errorf("expected input_tokens 10, got %d", result.Usage.InputTokens)
+	}
+	if result.Usage.OutputTokens != 5 {
+		t.Errorf("expected output_tokens 5, got %d", result.Usage.OutputTokens)
+	}
+}
+
+// TestAnthropicWriter_ErrorFromRoutes tests error handling when routes.go sends
+// gin.H{"error": "message"} without a StatusCode field (which is the common case)
+func TestAnthropicWriter_ErrorFromRoutes(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	tests := []struct {
+		name          string
+		statusCode    int
+		errorPayload  any
+		wantErrorType string
+		wantMessage   string
+	}{
+		// routes.go sends errors without StatusCode in JSON, so we must use HTTP status
+		{
+			name:          "404 with gin.H error (model not found)",
+			statusCode:    http.StatusNotFound,
+			errorPayload:  gin.H{"error": "model 'nonexistent' not found"},
+			wantErrorType: "not_found_error",
+			wantMessage:   "model 'nonexistent' not found",
+		},
+		{
+			name:          "400 with gin.H error (bad request)",
+			statusCode:    http.StatusBadRequest,
+			errorPayload:  gin.H{"error": "model is required"},
+			wantErrorType: "invalid_request_error",
+			wantMessage:   "model is required",
+		},
+		{
+			name:          "500 with gin.H error (internal error)",
+			statusCode:    http.StatusInternalServerError,
+			errorPayload:  gin.H{"error": "something went wrong"},
+			wantErrorType: "api_error",
+			wantMessage:   "something went wrong",
+		},
+		{
+			name:       "404 with api.StatusError",
+			statusCode: http.StatusNotFound,
+			errorPayload: api.StatusError{
+				StatusCode:   http.StatusNotFound,
+				ErrorMessage: "model not found via StatusError",
+			},
+			wantErrorType: "not_found_error",
+			wantMessage:   "model not found via StatusError",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			router := gin.New()
+			router.Use(AnthropicMessagesMiddleware())
+			router.POST("/v1/messages", func(c *gin.Context) {
+				// Simulate what routes.go does - set status and write error JSON
+				data, _ := json.Marshal(tt.errorPayload)
+				c.Writer.WriteHeader(tt.statusCode)
+				_, _ = c.Writer.Write(data)
+			})
+
+			body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
+			req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+			req.Header.Set("Content-Type", "application/json")
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if resp.Code != tt.statusCode {
+				t.Errorf("expected status %d, got %d", tt.statusCode, resp.Code)
+			}
+
+			var errResp anthropic.ErrorResponse
+			if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+				t.Fatalf("failed to unmarshal error response: %v\nbody: %s", err, resp.Body.String())
+			}
+
+			if errResp.Type != "error" {
+				t.Errorf("expected type 'error', got %q", errResp.Type)
+			}
+			if errResp.Error.Type != tt.wantErrorType {
+				t.Errorf("expected error type %q, got %q", tt.wantErrorType, errResp.Error.Type)
+			}
+			if errResp.Error.Message != tt.wantMessage {
+				t.Errorf("expected message %q, got %q", tt.wantMessage, errResp.Error.Message)
+			}
+		})
+	}
+}
+
+func TestAnthropicMessagesMiddleware_SetsRelaxThinkingFlag(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	var flagSet bool
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware())
+	router.POST("/v1/messages", func(c *gin.Context) {
+		_, flagSet = c.Get("relax_thinking")
+		c.Status(http.StatusOK)
+	})
+
+	body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
+	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if !flagSet {
+		t.Error("expected relax_thinking flag to be set in context")
+	}
+}

middleware/openai.go

@@ -8,6 +8,7 @@ import (
 	"math/rand"
 	"net/http"
 	"strings"
+	"time"
 
 	"github.com/gin-gonic/gin"
 
@@ -441,6 +442,7 @@ type ResponsesWriter struct {
 	stream     bool
 	responseID string
 	itemID     string
+	request    openai.ResponsesRequest
 }
 
 func (w *ResponsesWriter) writeEvent(eventType string, data any) error {
@@ -478,7 +480,9 @@ func (w *ResponsesWriter) writeResponse(data []byte) (int, error) {
 
 	// Non-streaming response
 	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	response := openai.ToResponse(w.model, w.responseID, w.itemID, chatResponse)
+	response := openai.ToResponse(w.model, w.responseID, w.itemID, chatResponse, w.request)
+	completedAt := time.Now().Unix()
+	response.CompletedAt = &completedAt
 	return len(data), json.NewEncoder(w.ResponseWriter).Encode(response)
 }
 
@@ -523,11 +527,12 @@ func ResponsesMiddleware() gin.HandlerFunc {
 
 		w := &ResponsesWriter{
 			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
-			converter:  openai.NewResponsesStreamConverter(responseID, itemID, req.Model),
+			converter:  openai.NewResponsesStreamConverter(responseID, itemID, req.Model, req),
 			model:      req.Model,
 			stream:     streamRequested,
 			responseID: responseID,
 			itemID:     itemID,
+			request:    req,
 		}
 
 		// Set headers based on streaming mode
@@ -541,3 +546,112 @@ func ResponsesMiddleware() gin.HandlerFunc {
 		c.Next()
 	}
 }
+
+type ImageWriter struct {
+	BaseWriter
+}
+
+func (w *ImageWriter) writeResponse(data []byte) (int, error) {
+	var generateResponse api.GenerateResponse
+	if err := json.Unmarshal(data, &generateResponse); err != nil {
+		return 0, err
+	}
+
+	// Only write response when done with image
+	if generateResponse.Done && generateResponse.Image != "" {
+		w.ResponseWriter.Header().Set("Content-Type", "application/json")
+		return len(data), json.NewEncoder(w.ResponseWriter).Encode(openai.ToImageGenerationResponse(generateResponse))
+	}
+
+	return len(data), nil
+}
+
+func (w *ImageWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func ImageGenerationsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.ImageGenerationRequest
+		if err := c.ShouldBindJSON(&req); err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Prompt == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "prompt is required"))
+			return
+		}
+
+		if req.Model == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "model is required"))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(openai.FromImageGenerationRequest(req)); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &ImageWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+		}
+
+		c.Writer = w
+		c.Next()
+	}
+}
+
+func ImageEditsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.ImageEditRequest
+		if err := c.ShouldBindJSON(&req); err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Prompt == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "prompt is required"))
+			return
+		}
+
+		if req.Model == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "model is required"))
+			return
+		}
+
+		if req.Image == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "image is required"))
+			return
+		}
+
+		genReq, err := openai.FromImageEditRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(genReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &ImageWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+		}
+
+		c.Writer = w
+		c.Next()
+	}
+}

middleware/openai_test.go

@@ -961,3 +961,280 @@ func TestRetrieveMiddleware(t *testing.T) {
 		}
 	}
 }
+
+func TestImageGenerationsMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.GenerateRequest
+		err  openai.ErrorResponse
+	}
+
+	var capturedRequest *api.GenerateRequest
+
+	testCases := []testCase{
+		{
+			name: "image generation basic",
+			body: `{
+				"model": "test-model",
+				"prompt": "a beautiful sunset"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "a beautiful sunset",
+			},
+		},
+		{
+			name: "image generation with size",
+			body: `{
+				"model": "test-model",
+				"prompt": "a beautiful sunset",
+				"size": "512x768"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "a beautiful sunset",
+				Width:  512,
+				Height: 768,
+			},
+		},
+		{
+			name: "image generation missing prompt",
+			body: `{
+				"model": "test-model"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "prompt is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+		{
+			name: "image generation missing model",
+			body: `{
+				"prompt": "a beautiful sunset"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "model is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(ImageGenerationsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/generate", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if tc.err.Error.Message != "" {
+				var errResp openai.ErrorResponse
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+				if diff := cmp.Diff(tc.err, errResp); diff != "" {
+					t.Fatalf("errors did not match:\n%s", diff)
+				}
+				return
+			}
+
+			if resp.Code != http.StatusOK {
+				t.Fatalf("expected status 200, got %d: %s", resp.Code, resp.Body.String())
+			}
+
+			if diff := cmp.Diff(&tc.req, capturedRequest); diff != "" {
+				t.Fatalf("requests did not match:\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestImageWriterResponse(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	// Test that ImageWriter transforms GenerateResponse to OpenAI format
+	endpoint := func(c *gin.Context) {
+		resp := api.GenerateResponse{
+			Model:     "test-model",
+			CreatedAt: time.Unix(1234567890, 0).UTC(),
+			Done:      true,
+			Image:     "dGVzdC1pbWFnZS1kYXRh", // base64 of "test-image-data"
+		}
+		data, _ := json.Marshal(resp)
+		c.Writer.Write(append(data, '\n'))
+	}
+
+	router := gin.New()
+	router.Use(ImageGenerationsMiddleware())
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	body := `{"model": "test-model", "prompt": "test"}`
+	req, _ := http.NewRequest(http.MethodPost, "/api/generate", strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", resp.Code, resp.Body.String())
+	}
+
+	var imageResp openai.ImageGenerationResponse
+	if err := json.Unmarshal(resp.Body.Bytes(), &imageResp); err != nil {
+		t.Fatalf("failed to unmarshal response: %v", err)
+	}
+
+	if imageResp.Created != 1234567890 {
+		t.Errorf("expected created 1234567890, got %d", imageResp.Created)
+	}
+
+	if len(imageResp.Data) != 1 {
+		t.Fatalf("expected 1 image, got %d", len(imageResp.Data))
+	}
+
+	if imageResp.Data[0].B64JSON != "dGVzdC1pbWFnZS1kYXRh" {
+		t.Errorf("expected image data 'dGVzdC1pbWFnZS1kYXRh', got %s", imageResp.Data[0].B64JSON)
+	}
+}
+
+func TestImageEditsMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.GenerateRequest
+		err  openai.ErrorResponse
+	}
+
+	var capturedRequest *api.GenerateRequest
+
+	// Base64-encoded test image (1x1 pixel PNG)
+	testImage := "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII="
+	decodedImage, _ := base64.StdEncoding.DecodeString("iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=")
+
+	testCases := []testCase{
+		{
+			name: "image edit basic",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue",
+				"image": "` + testImage + `"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "make it blue",
+				Images: []api.ImageData{decodedImage},
+			},
+		},
+		{
+			name: "image edit with size",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue",
+				"image": "` + testImage + `",
+				"size": "512x768"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "make it blue",
+				Images: []api.ImageData{decodedImage},
+				Width:  512,
+				Height: 768,
+			},
+		},
+		{
+			name: "image edit missing prompt",
+			body: `{
+				"model": "test-model",
+				"image": "` + testImage + `"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "prompt is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+		{
+			name: "image edit missing model",
+			body: `{
+				"prompt": "make it blue",
+				"image": "` + testImage + `"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "model is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+		{
+			name: "image edit missing image",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "image is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(ImageEditsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/generate", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if tc.err.Error.Message != "" {
+				var errResp openai.ErrorResponse
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+				if diff := cmp.Diff(tc.err, errResp); diff != "" {
+					t.Fatalf("errors did not match:\n%s", diff)
+				}
+				return
+			}
+
+			if resp.Code != http.StatusOK {
+				t.Fatalf("expected status 200, got %d: %s", resp.Code, resp.Body.String())
+			}
+
+			if diff := cmp.Diff(&tc.req, capturedRequest); diff != "" {
+				t.Fatalf("requests did not match:\n%s", diff)
+			}
+		})
+	}
+}

ml/backend.go

@@ -162,6 +162,7 @@ type Tensor interface {
 	AvgPool2D(ctx Context, k, s int, p float32) Tensor
 	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 	Conv3D(ctx Context, weight Tensor, c, s0, s1, s2, p0, p1, p2, d0, d1, d2 int) Tensor
+	SSMConv(ctx Context, kernel Tensor) Tensor
 
 	IM2Col(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 

ml/backend/ggml/ggml.go

@@ -1641,6 +1641,13 @@ func (t *Tensor) Conv3D(ctx ml.Context, t2 ml.Tensor, c, s0, s1, s2, p0, p1, p2,
 	return tt
 }
 
+func (t *Tensor) SSMConv(ctx ml.Context, kernel ml.Tensor) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_ssm_conv(ctx.(*Context).ctx, t.t, kernel.(*Tensor).t),
+	}
+}
+
 func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
 	return &Tensor{
 		b: t.b,

model/models/glm4moelite/model.go

@@ -0,0 +1,304 @@
+package glm4moelite
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Options struct {
+	numExpertsUsed      int
+	numExperts          int
+	normTopKProb        bool
+	routedScalingFactor float32
+
+	kvLoraRank,
+	qkNopeHeadDim,
+	qkRopeHeadDim,
+	kqNopeHeadDim,
+	qkHeadDim int
+	qLoraRank int
+	vHeadDim  int
+
+	hiddenSize,
+	numHeads,
+	numKVHeads int
+
+	eps,
+	ropeBase float32
+	kqScale float64
+}
+
+func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, t, p ml.Tensor) ml.Tensor {
+	return nn.RoPE(ctx, t, p, o.qkRopeHeadDim, o.ropeBase, 1.0)
+}
+
+type Attention struct {
+	Q *nn.Linear `gguf:"attn_q"`
+
+	QA     *nn.Linear  `gguf:"attn_q_a"`
+	QANorm *nn.RMSNorm `gguf:"attn_q_a_norm"`
+	QB     *nn.Linear  `gguf:"attn_q_b"`
+
+	KVA     *nn.Linear  `gguf:"attn_kv_a_mqa"`
+	KVANorm *nn.RMSNorm `gguf:"attn_kv_a_norm"`
+	KVB     *nn.Linear  `gguf:"attn_kv_b"`
+
+	Output *nn.Linear `gguf:"attn_out,alt:attn_output"`
+}
+
+func (attn *Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	seqLength := hiddenStates.Dim(1)
+
+	var query ml.Tensor
+	if opts.qLoraRank == 0 {
+		query = attn.Q.Forward(ctx, hiddenStates)
+	} else {
+		query = attn.QA.Forward(ctx, hiddenStates)
+		query = attn.QANorm.Forward(ctx, query, opts.eps)
+		query = attn.QB.Forward(ctx, query)
+	}
+
+	query = query.Reshape(ctx, query.Dim(0)/opts.numHeads, opts.numHeads, seqLength)
+	queryChunks := query.ChunkSections(ctx, 0, opts.qkNopeHeadDim, opts.qkRopeHeadDim)
+
+	compressedKV := attn.KVA.Forward(ctx, hiddenStates)
+	kPass := compressedKV.Slice(ctx, 0, 0, opts.kvLoraRank, 1)
+	kRot := compressedKV.View(ctx,
+		opts.kvLoraRank*compressedKV.Stride(0), opts.qkRopeHeadDim,
+		compressedKV.Stride(1), 1,
+		compressedKV.Stride(1), compressedKV.Dim(1),
+	)
+
+	qRot := opts.applyRotaryPositionEmbeddings(ctx, queryChunks[1], positions)
+	kRot = opts.applyRotaryPositionEmbeddings(ctx, kRot, positions)
+	kPass = attn.KVANorm.Forward(ctx, kPass, opts.eps)
+	kPass = attn.KVB.Forward(ctx, kPass)
+
+	kv := kPass.Reshape(ctx, kPass.Dim(0)/opts.numKVHeads, opts.numKVHeads, seqLength)
+	kvChunks := kv.ChunkSections(ctx, 0, opts.kqNopeHeadDim, opts.vHeadDim)
+
+	kRot = kRot.Repeat(ctx, 1, queryChunks[0].Dim(1))
+	query = qRot.Concat(ctx, queryChunks[0], 0)
+	key := kRot.Concat(ctx, kvChunks[0], 0)
+	attention := nn.Attention(ctx, query, key, kvChunks[1], opts.kqScale, cache)
+
+	attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), seqLength)
+	return attn.Output.Forward(ctx, attention)
+}
+
+type MLP interface {
+	Forward(ml.Context, ml.Tensor, *Options) ml.Tensor
+}
+
+type sparse struct {
+	Router       *nn.Linear `gguf:"ffn_gate_inp"`
+	Gate         *nn.Linear `gguf:"ffn_gate_exps"`
+	Up           *nn.Linear `gguf:"ffn_up_exps"`
+	Down         *nn.Linear `gguf:"ffn_down_exps"`
+	SharedExpert *dense     `gguf:",suf:_shexp"`
+	ExpProbsBias ml.Tensor  `gguf:"exp_probs_b.bias,alt:exp_probs_b"`
+}
+
+func (moe *sparse) Moe(ctx ml.Context, hiddenStates, topKIndices, topKWeights ml.Tensor, opts *Options) ml.Tensor {
+	hiddenStates = hiddenStates.Reshape(ctx, hiddenStates.Dim(0), 1, hiddenStates.Dim(1))
+
+	upStates := moe.Up.Weight.MulmatID(ctx, hiddenStates, topKIndices)
+	hiddenStates = moe.Gate.Weight.MulmatID(ctx, hiddenStates, topKIndices)
+	hiddenStates = hiddenStates.SILU(ctx, upStates)
+
+	experts := moe.Down.Weight.MulmatID(ctx, hiddenStates, topKIndices)
+	experts = experts.Mul(ctx, topKWeights)
+
+	nextStates := experts.View(ctx, 0, experts.Dim(0), experts.Stride(2), experts.Dim(2))
+	for i := 1; i < opts.numExpertsUsed; i++ {
+		nextStates = nextStates.Add(ctx, experts.View(ctx, i*experts.Stride(1), experts.Dim(0), experts.Stride(2), experts.Dim(2)))
+	}
+	return nextStates
+}
+
+func (moe *sparse) topKIndices(ctx ml.Context, scores ml.Tensor, opts *Options) ml.Tensor {
+	if moe.ExpProbsBias != nil {
+		scores = scores.Add(ctx, moe.ExpProbsBias)
+	}
+	topKIndices := scores.TopK(ctx, opts.numExpertsUsed)
+	return topKIndices
+}
+
+func (moe *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
+	residuals := hiddenStates
+
+	routerLogits := moe.Router.Forward(ctx, hiddenStates)
+	scores := routerLogits.Sigmoid(ctx)
+	topKIndices := moe.topKIndices(ctx, scores, opts)
+	topKWeights := scores.Reshape(ctx, 1, opts.numExperts, hiddenStates.Dim(1)).Rows(ctx, topKIndices)
+
+	if opts.normTopKProb {
+		topKWeights = topKWeights.Reshape(ctx, opts.numExpertsUsed, hiddenStates.Dim(1))
+		topKWeights = topKWeights.Div(ctx, topKWeights.SumRows(ctx))
+		topKWeights = topKWeights.Reshape(ctx, 1, opts.numExpertsUsed, hiddenStates.Dim(1))
+	}
+
+	topKWeights = topKWeights.Scale(ctx, float64(opts.routedScalingFactor))
+	hiddenStates = moe.Moe(ctx, hiddenStates, topKIndices, topKWeights, opts)
+	sharedExpertResult := moe.SharedExpert.Forward(ctx, residuals, opts)
+
+	hiddenStates = hiddenStates.Add(ctx, sharedExpertResult)
+	return hiddenStates
+}
+
+type dense struct {
+	Gate *nn.Linear `gguf:"ffn_gate"`
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+}
+
+func (mlp *dense) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	return mlp.Down.Forward(ctx, hiddenStates)
+}
+
+type Layer struct {
+	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
+	Attention     *Attention
+
+	MLPNorm *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP     MLP
+}
+
+func (t *Layer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	residual := hiddenStates
+	hiddenStates = t.AttentionNorm.Forward(ctx, hiddenStates, opts.eps)
+	hiddenStates = t.Attention.Forward(ctx, hiddenStates, positions, cache, opts)
+
+	if outputs != nil {
+		hiddenStates = hiddenStates.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+
+	hiddenStates = hiddenStates.Add(ctx, residual)
+	residual = hiddenStates
+
+	hiddenStates = t.MLPNorm.Forward(ctx, hiddenStates, opts.eps)
+	hiddenStates = t.MLP.Forward(ctx, hiddenStates, opts)
+	hiddenStates = hiddenStates.Add(ctx, residual)
+	return hiddenStates
+}
+
+type Model struct {
+	model.Base
+	model.BytePairEncoding
+
+	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
+	Layers         []Layer       `gguf:"blk"`
+
+	OutputNorm *nn.RMSNorm `gguf:"output_norm"`
+	Output     *nn.Linear  `gguf:"output,alt:token_embd"`
+
+	*Options
+}
+
+func New(c fs.Config) (model.Model, error) {
+	layers := make([]Layer, c.Uint("block_count"))
+
+	firstDenseLayerIndex := int(c.Uint("leading_dense_block_count"))
+	for i := range layers {
+		if i < firstDenseLayerIndex {
+			layers[i].MLP = &dense{}
+		} else {
+			layers[i].MLP = &sparse{}
+		}
+	}
+
+	keyLength := int(c.Uint("attention.key_length"))
+	valueLength := int(c.Uint("attention.value_length"))
+
+	kqScale := 1.0 / math.Sqrt(float64(keyLength))
+
+	var pre []string
+	switch c.String("tokenizer.ggml.pre") {
+	case "glm4":
+		pre = []string{
+			`(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`,
+		}
+	default:
+		return nil, model.ErrUnsupportedTokenizer
+	}
+
+	m := Model{
+		BytePairEncoding: model.NewBytePairEncoding(
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Types:  c.Ints("tokenizer.ggml.token_type"),
+				Merges: c.Strings("tokenizer.ggml.merges"),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
+			},
+			pre...,
+		),
+		Layers: layers,
+		Options: &Options{
+			hiddenSize:     int(c.Uint("embedding_length")),
+			numHeads:       int(c.Uint("attention.head_count")),
+			numKVHeads:     int(c.Uint("attention.head_count_kv")),
+			eps:            c.Float("attention.layer_norm_rms_epsilon"),
+			ropeBase:       c.Float("rope.freq_base"),
+			numExperts:     int(c.Uint("expert_count")),
+			numExpertsUsed: int(c.Uint("expert_used_count")),
+			normTopKProb:   c.Bool("expert_weights_norm", true),
+
+			qLoraRank:     int(c.Uint("attention.q_lora_rank")),
+			kvLoraRank:    int(c.Uint("attention.kv_lora_rank")),
+			qkHeadDim:     keyLength,
+			vHeadDim:      valueLength,
+			qkRopeHeadDim: int(c.Uint("rope.dimension_count")),
+			qkNopeHeadDim: keyLength - int(c.Uint("rope.dimension_count")),
+			kqNopeHeadDim: keyLength - int(c.Uint("rope.dimension_count")),
+
+			routedScalingFactor: c.Float("expert_weights_scale"),
+
+			kqScale: kqScale,
+		},
+	}
+
+	m.Cache = kvcache.NewCausalCache(m.Shift)
+	return &m, nil
+}
+
+func (m Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	return m.applyRotaryPositionEmbeddings(ctx, key, shift), nil
+}
+
+func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+	positions := ctx.Input().FromInts(batch.Positions, len(batch.Positions))
+
+	hiddenStates := m.TokenEmbedding.Forward(ctx, batch.Inputs)
+
+	for i, layer := range m.Layers {
+		m.Cache.SetLayer(i)
+
+		var outputs ml.Tensor
+		if i == len(m.Layers)-1 {
+			outputs = batch.Outputs
+		}
+
+		hiddenStates = layer.Forward(ctx, hiddenStates, positions, outputs, m.Cache, m.Options)
+	}
+
+	hiddenStates = m.OutputNorm.Forward(ctx, hiddenStates, m.eps)
+	return m.Output.Forward(ctx, hiddenStates), nil
+}
+
+func init() {
+	model.Register("glm4moelite", New)
+}

model/models/lfm2/cache.go

@@ -0,0 +1,410 @@
+package lfm2
+
+import (
+	"slices"
+
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/input"
+)
+
+var _ kvcache.Cache = (*HybridCache)(nil)
+
+// HybridCache stores:
+// - a standard causal KV cache for attention layers
+// - a per-sequence recurrent conv state for shortconv layers
+//
+// Conv state shape (per layer, per sequence): [dConv, hiddenSize] where dConv = L_cache - 1.
+// Stored internally as a tensor of shape [dConv * hiddenSize, maxSlots].
+type HybridCache struct {
+	kv *kvcache.Causal
+
+	backend      ml.Backend
+	dtype        ml.DType
+	maxSequences int
+
+	hiddenSize int
+	dConv      int
+
+	// slot mapping for recurrent state
+	slotForSeq map[int]int
+	refCount   []int
+	freeSlots  []int
+
+	// per-layer conv state buffers (allocated lazily)
+	convCtxs   map[int]ml.Context
+	convStates map[int]ml.Tensor // [dConv*hiddenSize, maxSlots]
+
+	// current forward batch (derived in StartForward)
+	curSeqs       []int
+	curSlots      []int
+	curSlotsInput ml.Tensor
+	curSeqTokens  int
+
+	// track if EnsureWritable has been called for this forward pass
+	writableEnsured bool
+	// track any error from EnsureWritable to propagate later
+	writableError error
+}
+
+func NewHybridCache(shift func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error), hiddenSize, dConv int) *HybridCache {
+	return &HybridCache{
+		kv:         kvcache.NewCausalCache(shift),
+		hiddenSize: hiddenSize,
+		dConv:      dConv,
+		slotForSeq: make(map[int]int),
+		convCtxs:   make(map[int]ml.Context),
+		convStates: make(map[int]ml.Tensor),
+	}
+}
+
+func (c *HybridCache) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int) {
+	c.backend = backend
+	c.dtype = dtype
+	c.maxSequences = maxSequences
+
+	// initialize slot allocator
+	c.refCount = make([]int, maxSequences)
+	c.freeSlots = c.freeSlots[:0]
+	for i := maxSequences - 1; i >= 0; i-- {
+		c.freeSlots = append(c.freeSlots, i)
+	}
+
+	c.kv.Init(backend, dtype, maxSequences, capacity, maxBatch)
+}
+
+func (c *HybridCache) Close() {
+	for _, ctx := range c.convCtxs {
+		ctx.Close()
+	}
+	c.kv.Close()
+}
+
+func (c *HybridCache) SetConfig(config ml.CacheConfig) {
+	c.kv.SetConfig(config)
+}
+
+func (c *HybridCache) SetLayer(layer int) {
+	c.kv.SetLayer(layer)
+}
+
+func (c *HybridCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
+	return c.kv.Get(ctx)
+}
+
+func (c *HybridCache) Put(ctx ml.Context, key, value ml.Tensor) {
+	c.kv.Put(ctx, key, value)
+}
+
+func (c *HybridCache) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
+	if err := c.kv.StartForward(ctx, batch, reserve); err != nil {
+		return err
+	}
+
+	// Derive equal-length sequence layout for shortconv.
+	// LFM2 shortconv assumes tokens form a [seq_tokens, seqs] grid.
+	seqCounts := make(map[int]int)
+	c.curSeqs = c.curSeqs[:0]
+	for _, s := range batch.Sequences {
+		if _, ok := seqCounts[s]; !ok {
+			c.curSeqs = append(c.curSeqs, s)
+		}
+		seqCounts[s]++
+	}
+
+	if len(c.curSeqs) == 0 {
+		return nil
+	}
+
+	nTokens := len(batch.Sequences)
+	nSeqs := len(c.curSeqs)
+	want := nTokens / nSeqs
+	for _, s := range c.curSeqs {
+		if seqCounts[s] != want {
+			return kvcache.ErrNotSupported
+		}
+	}
+
+	c.curSeqTokens = want
+
+	// When reserving memory for estimation, use fake slot assignments
+	// without modifying permanent state (slotForSeq, refCount)
+	if reserve {
+		c.curSlots = c.curSlots[:0]
+		slots := make([]int32, nSeqs)
+		for i := range nSeqs {
+			c.curSlots = append(c.curSlots, i)
+			slots[i] = int32(i)
+		}
+		c.curSlotsInput = ctx.Input().FromInts(slots, len(slots))
+		return nil
+	}
+
+	// Ensure slots exist for sequences in this batch
+	c.curSlots = c.curSlots[:0]
+	var newSlots []int // track newly allocated slots that need zeroing
+	for _, s := range c.curSeqs {
+		slot, ok := c.slotForSeq[s]
+		if !ok {
+			var err error
+			slot, err = c.allocSlot()
+			if err != nil {
+				return err
+			}
+			c.slotForSeq[s] = slot
+			c.refCount[slot] = 1
+			newSlots = append(newSlots, slot)
+		}
+		c.curSlots = append(c.curSlots, slot)
+	}
+
+	// Zero conv state for newly allocated slots to clear stale data from previous sequences
+	if len(newSlots) > 0 {
+		c.zeroConvSlots(ctx, newSlots)
+	}
+
+	// Create a tensor for the current slots
+	slots := make([]int32, len(c.curSlots))
+	for i, v := range c.curSlots {
+		slots[i] = int32(v)
+	}
+	c.curSlotsInput = ctx.Input().FromInts(slots, len(slots))
+
+	// Reset writable state for new forward pass
+	c.writableEnsured = false
+	c.writableError = nil
+
+	return nil
+}
+
+func (c *HybridCache) allocSlot() (int, error) {
+	if len(c.freeSlots) == 0 {
+		return 0, kvcache.ErrKvCacheFull
+	}
+	slot := c.freeSlots[len(c.freeSlots)-1]
+	c.freeSlots = c.freeSlots[:len(c.freeSlots)-1]
+	return slot, nil
+}
+
+func (c *HybridCache) freeSlot(slot int) {
+	// Bounds check before freeing
+	if slot >= 0 && slot < c.maxSequences {
+		c.freeSlots = append(c.freeSlots, slot)
+	}
+}
+
+// zeroConvSlots zeros the conv state for the given slots across all layers.
+// This must be called when recycling slots to prevent stale state from affecting new sequences.
+func (c *HybridCache) zeroConvSlots(ctx ml.Context, slots []int) {
+	if len(slots) == 0 || len(c.convStates) == 0 {
+		return
+	}
+
+	// Use input context for creating tensors
+	inputCtx := ctx.Input()
+
+	// Create slot indices tensor
+	slotIndices := make([]int32, len(slots))
+	for i, s := range slots {
+		slotIndices[i] = int32(s)
+	}
+	slotsTensor := inputCtx.FromInts(slotIndices, len(slotIndices))
+
+	// Create zero tensor for the slots (SetRows requires F32 source)
+	zeros := inputCtx.Zeros(ml.DTypeF32, c.dConv*c.hiddenSize, len(slots))
+
+	// Zero each layer's conv state for these slots
+	for _, buf := range c.convStates {
+		ctx.Forward(buf.SetRows(ctx, zeros, slotsTensor))
+	}
+}
+
+// EnsureWritable ensures that sequences in the current batch have private (non-shared) conv slots.
+// Returns an error if slot allocation fails.
+func (c *HybridCache) EnsureWritable(ctx ml.Context) error {
+	for i, seq := range c.curSeqs {
+		slot, ok := c.slotForSeq[seq]
+		if !ok {
+			continue
+		}
+
+		// Bounds check
+		if slot < 0 || slot >= len(c.refCount) {
+			continue
+		}
+
+		if c.refCount[slot] <= 1 {
+			continue
+		}
+
+		newSlot, err := c.allocSlot()
+		if err != nil {
+			return err
+		}
+		c.refCount[slot]--
+		c.refCount[newSlot] = 1
+		c.slotForSeq[seq] = newSlot
+		c.curSlots[i] = newSlot
+
+		// Copy existing conv state for all initialized layers
+		for _, buf := range c.convStates {
+			// buf: [dConv*hiddenSize, maxSlots]
+			src := buf.Rows(ctx, ctx.Input().FromInts([]int32{int32(slot)}, 1))
+			// SetRows requires F32 source
+			srcF32 := src.Cast(ctx, ml.DTypeF32)
+			ctx.Forward(buf.SetRows(ctx, srcF32, ctx.Input().FromInts([]int32{int32(newSlot)}, 1)))
+		}
+	}
+
+	// Rebuild current slots tensor
+	slots := make([]int32, len(c.curSlots))
+	for i, v := range c.curSlots {
+		slots[i] = int32(v)
+	}
+	c.curSlotsInput = ctx.Input().FromInts(slots, len(slots))
+
+	return nil
+}
+
+func (c *HybridCache) CopyPrefix(srcSeq, dstSeq int, prefixLen int32) {
+	// KV cache shares prefix metadata (no copy) which is correct for prefix reuse.
+	c.kv.CopyPrefix(srcSeq, dstSeq, prefixLen)
+
+	// For shortconv state we implement copy-on-write: dst shares the same slot as src.
+	// On the first write to dst, EnsureWritable will create a private slot.
+	if dstSlot, ok := c.slotForSeq[dstSeq]; ok {
+		// Bounds check before decrementing
+		if dstSlot >= 0 && dstSlot < len(c.refCount) {
+			c.refCount[dstSlot]--
+			if c.refCount[dstSlot] <= 0 {
+				c.refCount[dstSlot] = 0
+				c.freeSlot(dstSlot)
+			}
+		}
+		delete(c.slotForSeq, dstSeq)
+	}
+
+	srcSlot, ok := c.slotForSeq[srcSeq]
+	if !ok {
+		// src may not have a slot yet; dst will allocate on demand
+		return
+	}
+
+	// Bounds check before incrementing
+	if srcSlot >= 0 && srcSlot < len(c.refCount) {
+		c.slotForSeq[dstSeq] = srcSlot
+		c.refCount[srcSlot]++
+	}
+}
+
+func (c *HybridCache) CanResume(seq int, pos int32) bool {
+	return c.kv.CanResume(seq, pos)
+}
+
+func (c *HybridCache) Remove(seq int, beginIndex, endIndex int32) error {
+	if err := c.kv.Remove(seq, beginIndex, endIndex); err != nil {
+		return err
+	}
+
+	// For recurrent state, any removal invalidates the state because
+	// the state at position N depends on all previous positions.
+	// Drop the slot mapping so it resets on next use.
+	slot, ok := c.slotForSeq[seq]
+	if !ok {
+		return nil
+	}
+
+	// Bounds check
+	if slot < 0 || slot >= len(c.refCount) {
+		delete(c.slotForSeq, seq)
+		return nil
+	}
+
+	c.refCount[slot]--
+	if c.refCount[slot] <= 0 {
+		c.refCount[slot] = 0
+		c.freeSlot(slot)
+	}
+	delete(c.slotForSeq, seq)
+
+	return nil
+}
+
+func (c *HybridCache) slotsTensor() ml.Tensor {
+	return c.curSlotsInput
+}
+
+func (c *HybridCache) seqTokens() int {
+	return c.curSeqTokens
+}
+
+func (c *HybridCache) numSeqs() int {
+	return len(c.curSeqs)
+}
+
+func (c *HybridCache) convBuffer(ctx ml.Context, layer int) ml.Tensor {
+	if buf, ok := c.convStates[layer]; ok {
+		return buf
+	}
+
+	if _, ok := c.convCtxs[layer]; !ok {
+		c.convCtxs[layer] = c.backend.NewContextSize(1).Layer(layer)
+	}
+
+	buf := c.convCtxs[layer].Zeros(c.dtype, c.dConv*c.hiddenSize, c.maxSequences)
+	c.convStates[layer] = buf
+	return buf
+}
+
+// ConvState returns the conv state for current batch sequences as shape [dConv, hiddenSize, nSeqs].
+// Returns an error if copy-on-write allocation fails.
+func (c *HybridCache) ConvState(ctx ml.Context, layer int) (ml.Tensor, error) {
+	if !c.writableEnsured {
+		needsWritable := false
+		for _, seq := range c.curSeqs {
+			slot, ok := c.slotForSeq[seq]
+			if !ok {
+				continue
+			}
+			if slot >= 0 && slot < len(c.refCount) && c.refCount[slot] > 1 {
+				needsWritable = true
+				break
+			}
+		}
+
+		if needsWritable {
+			if err := c.EnsureWritable(ctx); err != nil {
+				c.writableError = err
+			}
+		}
+		c.writableEnsured = true
+	}
+
+	if c.writableError != nil {
+		return nil, c.writableError
+	}
+
+	buf := c.convBuffer(ctx, layer)
+	cur := buf.Rows(ctx, c.slotsTensor())
+	return cur.Reshape(ctx, c.dConv, c.hiddenSize, c.numSeqs()), nil
+}
+
+// UpdateConvState writes a new conv state for current batch sequences.
+// newState must have shape [dConv, hiddenSize, nSeqs].
+func (c *HybridCache) UpdateConvState(ctx ml.Context, layer int, newState ml.Tensor) {
+	buf := c.convBuffer(ctx, layer)
+	src := newState.Reshape(ctx, c.dConv*c.hiddenSize, c.numSeqs())
+	// SetRows requires F32 source
+	srcF32 := src.Cast(ctx, ml.DTypeF32)
+	ctx.Forward(buf.SetRows(ctx, srcF32, c.slotsTensor()))
+}
+
+// IsSupportedForBatch returns true if the current batch layout supports shortconv.
+func (c *HybridCache) IsSupportedForBatch() bool {
+	return c.curSeqTokens > 0 && len(c.curSeqs) > 0
+}
+
+// Seqs returns the ordered unique sequences for the current forward pass.
+func (c *HybridCache) Seqs() []int {
+	return slices.Clone(c.curSeqs)
+}