server: add tests and fix isHuggingFaceURL edge case

- Add comprehensive tests for isHuggingFaceURL and getNumDownloadParts
- Fix bug where domains ending in huggingface.co (like nothuggingface.co)
  would incorrectly match as HuggingFace URLs
- Improve code comments with more detailed documentation

CMakeLists.txt

@@ -190,7 +190,7 @@ if(MLX_ENGINE)
     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_INCLUDE_REGEXES cublas cublasLt cudart nvrtc cudnn nccl
             PRE_EXCLUDE_REGEXES ".*"
         RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
         LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX

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 git \
+    && dnf install -y clang ccache \
     && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
 ENV CC=clang CXX=clang++
 
@@ -149,7 +149,6 @@ 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
@@ -157,6 +156,14 @@ 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}
+COPY . .
+ARG GOFLAGS="'-ldflags=-w -s'"
+ENV CGO_ENABLED=1
+ARG CGO_CFLAGS
+ARG CGO_CXXFLAGS
+RUN mkdir -p dist/bin
+RUN --mount=type=cache,target=/root/.cache/go-build \
+    go build -tags mlx -trimpath -buildmode=pie -o dist/bin/ollama-mlx .
 
 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -165,14 +172,12 @@ 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
-ENV CGO_CFLAGS="-I/go/src/github.com/ollama/ollama/build/_deps/mlx-c-src"
+ARG CGO_CFLAGS
 ARG CGO_CXXFLAGS
 RUN --mount=type=cache,target=/root/.cache/go-build \
-    go build -tags mlx -trimpath -buildmode=pie -o /bin/ollama .
+    go build -trimpath -buildmode=pie -o /bin/ollama .
 
 FROM --platform=linux/amd64 scratch AS amd64
 # COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
@@ -180,6 +185,7 @@ 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/
+COPY --from=mlx     /go/src/github.com/ollama/ollama/dist/bin/ /bin/
 
 FROM --platform=linux/arm64 scratch AS arm64
 # COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/

MLX_VERSION

@@ -1 +0,0 @@
-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.
@@ -260,38 +260,6 @@ 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.
@@ -322,7 +290,6 @@ 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)
@@ -454,7 +421,7 @@ 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.)
@@ -526,7 +493,7 @@ 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)
@@ -669,7 +636,6 @@ 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.
@@ -678,5 +644,4 @@ 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)

api/types.go

@@ -127,20 +127,6 @@ 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].
@@ -749,7 +735,7 @@ 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"`
+	ModelInfo     map[string]any     `json:"model_info,omitempty"`
 	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
 	Tensors       []Tensor           `json:"tensors,omitempty"`
 	Capabilities  []model.Capability `json:"capabilities,omitempty"`
@@ -874,20 +860,6 @@ 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

@@ -14,7 +14,6 @@ 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
@@ -41,13 +40,6 @@ 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"]) {
@@ -286,18 +278,7 @@ 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

@@ -35,7 +35,6 @@ 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"
@@ -47,9 +46,8 @@ 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"
+	imagegenclient "github.com/ollama/ollama/x/imagegen/client"
 )
 
 const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
@@ -95,87 +93,15 @@ 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(".") {
+			if imagegen.IsTensorModelDir(".") {
 				quantize, _ := cmd.Flags().GetString("quantize")
-				return xcreateclient.CreateModel(xcreateclient.CreateOptions{
-					ModelName: modelName,
-					ModelDir:  ".",
-					Quantize:  quantize,
-				}, p)
+				return imagegenclient.CreateModel(args[0], ".", quantize, p)
 			}
 			reader = strings.NewReader("FROM .\n")
 		} else {
@@ -208,7 +134,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	}
 	spinner.Stop()
 
-	req.Model = modelName
+	req.Model = args[0]
 	quantize, _ := cmd.Flags().GetString("quantize")
 	if quantize != "" {
 		req.Quantize = quantize
@@ -601,7 +527,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	// Check if this is an image generation model
-	if slices.Contains(info.Capabilities, model.CapabilityImage) {
+	if slices.Contains(info.Capabilities, model.CapabilityImageGeneration) {
 		if opts.Prompt == "" && !interactive {
 			return errors.New("image generation models require a prompt. Usage: ollama run " + name + " \"your prompt here\"")
 		}
@@ -900,11 +826,11 @@ 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:     arg,
+			Model:     args[0],
 			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", arg)
+				fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", args[0])
 			}
 		}
 
@@ -1816,22 +1742,15 @@ 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: 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,
+		Use:     "create MODEL",
+		Short:   "Create a model",
+		Args:    cobra.ExactArgs(1),
+		PreRunE: checkServerHeartbeat,
+		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",
@@ -1986,7 +1905,6 @@ 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{
@@ -2027,7 +1945,6 @@ func NewCLI() *cobra.Command {
 		copyCmd,
 		deleteCmd,
 		runnerCmd,
-		config.ConfigCmd(checkServerHeartbeat),
 	)
 
 	return rootCmd

cmd/cmd_test.go

@@ -1555,7 +1555,7 @@ func TestShowInfoImageGen(t *testing.T) {
 			ParameterSize:     "10.3B",
 			QuantizationLevel: "FP8",
 		},
-		Capabilities: []model.Capability{model.CapabilityImage},
+		Capabilities: []model.Capability{model.CapabilityImageGeneration},
 		Requires:     "0.14.0",
 	}, false, &b)
 	if err != nil {

cmd/config/claude.go

@@ -1,36 +0,0 @@
-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

@@ -1,42 +0,0 @@
-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

@@ -1,61 +0,0 @@
-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

@@ -1,28 +0,0 @@
-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

@@ -1,115 +0,0 @@
-// 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

@@ -1,373 +0,0 @@
-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

@@ -1,164 +0,0 @@
-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

@@ -1,454 +0,0 @@
-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

@@ -1,99 +0,0 @@
-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

@@ -1,502 +0,0 @@
-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

@@ -1,361 +0,0 @@
-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

@@ -1,188 +0,0 @@
-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

@@ -1,203 +0,0 @@
-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

@@ -1,437 +0,0 @@
-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

@@ -1,499 +0,0 @@
-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

@@ -1,913 +0,0 @@
-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: "Press Enter to send",
+		AltPlaceholder: `Use """ to end multi-line input`,
 	})
 	if err != nil {
 		return err
@@ -159,7 +159,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			sb.WriteString(before)
 			if !ok {
 				fmt.Fprintln(&sb)
-				scanner.Prompt.UseAlt = true
 				continue
 			}
 

convert/convert.go

@@ -311,10 +311,6 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 		conv = &deepseekocr{}
 	case "DeepseekV3ForCausalLM":
 		conv = &deepseek2Model{}
-	case "Glm4MoeLiteForCausalLM":
-		conv = &glm4MoeLiteModel{}
-	case "Lfm2ForCausalLM":
-		conv = &lfm2Model{}
 	default:
 		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}

convert/convert_glm4moelite.go

@@ -1,150 +0,0 @@
-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_lfm2.go

@@ -1,100 +0,0 @@
-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/reader.go

@@ -40,7 +40,6 @@ 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/api.md

@@ -16,7 +16,6 @@
 - [Generate Embeddings](#generate-embeddings)
 - [List Running Models](#list-running-models)
 - [Version](#version)
-- [Experimental: Image Generation](#image-generation-experimental)
 
 ## Conventions
 
@@ -59,15 +58,6 @@ 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.
@@ -1877,55 +1867,3 @@ 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

@@ -21,7 +21,6 @@ ollama pull glm-4.7:cloud
 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
 ```
@@ -248,13 +247,12 @@ curl -X POST http://localhost:11434/v1/messages \
 [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
+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
 ```

docs/api/openai-compatibility.mdx

@@ -275,73 +275,6 @@ 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("what is ollama?");
+const results = await client.webSearch({ query: "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("https://ollama.com");
+const fetchResult = await client.webFetch({ url: "https://ollama.com" });
 console.log(JSON.stringify(fetchResult, null, 2));
 ```
 

docs/docs.json

@@ -111,9 +111,7 @@
               "/integrations/zed",
               "/integrations/roo-code",
               "/integrations/n8n",
-              "/integrations/xcode",
-              "/integrations/onyx",
-              "/integrations/marimo"
+              "/integrations/xcode"
             ]
           },
           {

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 4096 tokens.
+By default, Ollama uses a context window size of 2048 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

@@ -2,12 +2,6 @@
 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):
@@ -31,24 +25,22 @@ 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
+export ANTHROPIC_API_KEY=ollama
 ```
 
 2. Run Claude Code with an Ollama model:
 
 ```shell
-claude --model gpt-oss:20b
+claude --model qwen3-coder
 ```
 
 Or run with environment variables inline:
 
 ```shell
-ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 claude --model gpt-oss:20b
+ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
 ```
 
-**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
@@ -75,4 +67,3 @@ claude --model glm-4.7:cloud
 ### 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

@@ -1,73 +0,0 @@
----
-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

@@ -1,63 +0,0 @@
----
-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

@@ -1,5 +1,5 @@
 ---
-title: Linux
+title: "Linux"
 ---
 
 ## Install
@@ -13,15 +13,14 @@ curl -fsSL https://ollama.com/install.sh | sh
 ## Manual install
 
 <Note>
-  If you are upgrading from a prior version, you should remove the old libraries
-  with `sudo rm -rf /usr/lib/ollama` first.
+  If you are upgrading from a prior version, you should remove the old libraries with `sudo rm -rf /usr/lib/ollama` first.
 </Note>
 
 Download and extract the package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \
-    | sudo tar x -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz \
+    | sudo tar zx -C /usr
 ```
 
 Start Ollama:
@@ -41,8 +40,8 @@ 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.tar.zst \
-    | sudo tar x -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz \
+    | sudo tar zx -C /usr
 ```
 
 ### ARM64 install
@@ -50,8 +49,8 @@ curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tar.zst \
 Download and extract the ARM64-specific package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-arm64.tar.zst \
-    | sudo tar x -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-arm64.tgz \
+    | sudo tar zx -C /usr
 ```
 
 ### Adding Ollama as a startup service (recommended)
@@ -113,11 +112,7 @@ sudo systemctl status ollama
 ```
 
 <Note>
-  While AMD has contributed the `amdgpu` driver upstream to the official linux
-  kernel source, the version is older and may not support all ROCm features. We
-  recommend you install the latest driver from
-  https://www.amd.com/en/support/linux-drivers for best support of your Radeon
-  GPU.
+  While AMD has contributed the `amdgpu` driver upstream to the official linux kernel source, the version is older and may not support all ROCm features. We recommend you install the latest driver from https://www.amd.com/en/support/linux-drivers for best support of your Radeon GPU.
 </Note>
 
 ## Customizing
@@ -146,8 +141,8 @@ curl -fsSL https://ollama.com/install.sh | sh
 Or by re-downloading Ollama:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \
-    | sudo tar x -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz \
+    | sudo tar zx -C /usr
 ```
 
 ## Installing specific versions
@@ -196,4 +191,4 @@ Remove the downloaded models and Ollama service user and group:
 sudo userdel ollama
 sudo groupdel ollama
 sudo rm -r /usr/share/ollama
-```
+```

fs/ggml/ggml.go

@@ -269,8 +269,6 @@ func (kv KV) OllamaEngineRequired() bool {
 		"qwen25vl",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
-		"glm4moelite",
-		"lfm2",
 	}, kv.Architecture())
 }
 
@@ -858,9 +856,7 @@ func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
 		"bert",
 		"gemma3",
-		"glm4moelite",
 		"gptoss", "gpt-oss",
-		"lfm2",
 		"mistral3",
 		"olmo3",
 		"qwen3", "qwen3moe",

integration/imagegen_test.go

@@ -1,148 +0,0 @@
-//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.Get("location"); !ok {
+				if _, ok := lastToolCall.Function.Arguments["location"]; !ok {
 					t.Errorf("expected tool arguments to include 'location', got: %s", lastToolCall.Function.Arguments.String())
 				}
 			case <-ctx.Done():

integration/utils_test.go

@@ -38,7 +38,6 @@ 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",
@@ -144,7 +143,6 @@ var (
 		"granite3.3",
 		"hermes3",
 		"internlm2",
-		"lfm2.5-thinking",
 		"llama-guard3",
 		"llama-pro",
 		"llama2-chinese",
@@ -265,7 +263,6 @@ var (
 		"snowflake-arctic-embed2",
 	}
 	libraryToolsModels = []string{
-		"lfm2.5-thinking",
 		"qwen3-vl",
 		"gpt-oss:20b",
 		"gpt-oss:120b",

llm/server.go

@@ -1464,12 +1464,6 @@ 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
@@ -1518,15 +1512,6 @@ 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/paths.go

@@ -1,95 +0,0 @@
-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/openai.go

@@ -8,7 +8,6 @@ import (
 	"math/rand"
 	"net/http"
 	"strings"
-	"time"
 
 	"github.com/gin-gonic/gin"
 
@@ -442,7 +441,6 @@ type ResponsesWriter struct {
 	stream     bool
 	responseID string
 	itemID     string
-	request    openai.ResponsesRequest
 }
 
 func (w *ResponsesWriter) writeEvent(eventType string, data any) error {
@@ -480,9 +478,7 @@ 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, w.request)
-	completedAt := time.Now().Unix()
-	response.CompletedAt = &completedAt
+	response := openai.ToResponse(w.model, w.responseID, w.itemID, chatResponse)
 	return len(data), json.NewEncoder(w.ResponseWriter).Encode(response)
 }
 
@@ -527,12 +523,11 @@ func ResponsesMiddleware() gin.HandlerFunc {
 
 		w := &ResponsesWriter{
 			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
-			converter:  openai.NewResponsesStreamConverter(responseID, itemID, req.Model, req),
+			converter:  openai.NewResponsesStreamConverter(responseID, itemID, req.Model),
 			model:      req.Model,
 			stream:     streamRequested,
 			responseID: responseID,
 			itemID:     itemID,
-			request:    req,
 		}
 
 		// Set headers based on streaming mode
@@ -546,112 +541,3 @@ 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

@@ -961,280 +961,3 @@ 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

@@ -162,7 +162,6 @@ 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

@@ -1641,13 +1641,6 @@ 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

@@ -1,304 +0,0 @@
-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

@@ -1,410 +0,0 @@
-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)
-}

model/models/lfm2/cache_test.go

@@ -1,444 +0,0 @@
-package lfm2
-
-import (
-	"testing"
-
-	"github.com/ollama/ollama/kvcache"
-	"github.com/ollama/ollama/ml"
-)
-
-// TestHybridCache tests verify the slot management logic of HybridCache.
-// These tests focus on the recurrent state slot allocation, reference counting,
-// and copy-on-write semantics without requiring a full ML backend.
-
-// createSlotOnlyCache creates a HybridCache with only the slot management
-// fields initialized. Used to test slot logic in isolation.
-func createSlotOnlyCache(maxSequences int) *HybridCache {
-	return &HybridCache{
-		hiddenSize:   256,
-		dConv:        3,
-		maxSequences: maxSequences,
-		refCount:     make([]int, maxSequences),
-		freeSlots:    initFreeSlots(maxSequences),
-		slotForSeq:   make(map[int]int),
-		convCtxs:     make(map[int]ml.Context),
-		convStates:   make(map[int]ml.Tensor),
-	}
-}
-
-func initFreeSlots(n int) []int {
-	slots := make([]int, 0, n)
-	for i := n - 1; i >= 0; i-- {
-		slots = append(slots, i)
-	}
-	return slots
-}
-
-func TestHybridCache_SlotAllocation(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Verify initial state
-	if len(cache.freeSlots) != 4 {
-		t.Errorf("expected 4 free slots, got %d", len(cache.freeSlots))
-	}
-
-	// Allocate all slots
-	for range 4 {
-		slot, err := cache.allocSlot()
-		if err != nil {
-			t.Fatalf("allocSlot failed: %v", err)
-		}
-		cache.refCount[slot] = 1
-	}
-
-	// Should be full now
-	if len(cache.freeSlots) != 0 {
-		t.Errorf("expected 0 free slots, got %d", len(cache.freeSlots))
-	}
-
-	// Trying to allocate another should fail
-	_, err := cache.allocSlot()
-	if err != kvcache.ErrKvCacheFull {
-		t.Errorf("expected ErrKvCacheFull, got %v", err)
-	}
-}
-
-func TestHybridCache_SlotReuse(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Allocate a slot
-	slot1, _ := cache.allocSlot()
-	cache.refCount[slot1] = 1
-
-	// Free it
-	cache.refCount[slot1] = 0
-	cache.freeSlot(slot1)
-
-	// Allocate again - should get the same slot back (LIFO)
-	slot2, _ := cache.allocSlot()
-	if slot2 != slot1 {
-		t.Errorf("expected slot %d to be reused, got %d", slot1, slot2)
-	}
-}
-
-func TestHybridCache_SlotRefCounting_ShareSlot(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Allocate slot for seq 1
-	slot1, _ := cache.allocSlot()
-	cache.slotForSeq[1] = slot1
-	cache.refCount[slot1] = 1
-
-	// Simulate sharing slot with seq 2 (copy-on-write style)
-	cache.slotForSeq[2] = slot1
-	cache.refCount[slot1]++
-
-	// Should share the same slot
-	if cache.slotForSeq[2] != slot1 {
-		t.Errorf("expected seq 2 to share slot %d, got %d", slot1, cache.slotForSeq[2])
-	}
-
-	// Ref count should be 2
-	if cache.refCount[slot1] != 2 {
-		t.Errorf("expected refCount 2, got %d", cache.refCount[slot1])
-	}
-}
-
-func TestHybridCache_SlotRefCounting_DecRef(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Allocate slot for seq 1
-	slot1, _ := cache.allocSlot()
-	cache.slotForSeq[1] = slot1
-	cache.refCount[slot1] = 1
-
-	// Share with seq 2
-	cache.slotForSeq[2] = slot1
-	cache.refCount[slot1]++
-
-	// Unshare seq 2
-	cache.refCount[slot1]--
-	delete(cache.slotForSeq, 2)
-
-	// Ref count should be back to 1
-	if cache.refCount[slot1] != 1 {
-		t.Errorf("expected refCount 1 after unshare, got %d", cache.refCount[slot1])
-	}
-
-	// Seq 2 should no longer have a slot
-	if _, ok := cache.slotForSeq[2]; ok {
-		t.Error("seq 2 should not have a slot after unshare")
-	}
-}
-
-func TestHybridCache_SlotFreeWhenUnused(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	initialFreeSlots := len(cache.freeSlots)
-
-	// Allocate slot for seq 1
-	slot1, _ := cache.allocSlot()
-	cache.slotForSeq[1] = slot1
-	cache.refCount[slot1] = 1
-
-	// Free the slot when refCount drops to 0
-	cache.refCount[slot1]--
-	if cache.refCount[slot1] <= 0 {
-		cache.refCount[slot1] = 0
-		cache.freeSlot(slot1)
-	}
-	delete(cache.slotForSeq, 1)
-
-	// Slot should be freed
-	if len(cache.freeSlots) != initialFreeSlots {
-		t.Errorf("expected %d free slots, got %d", initialFreeSlots, len(cache.freeSlots))
-	}
-
-	// Ref count should be 0
-	if cache.refCount[slot1] != 0 {
-		t.Errorf("expected refCount 0, got %d", cache.refCount[slot1])
-	}
-}
-
-func TestHybridCache_SlotOverwrite(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Allocate slots for seq 1 and seq 2
-	slot1, _ := cache.allocSlot()
-	cache.slotForSeq[1] = slot1
-	cache.refCount[slot1] = 1
-
-	slot2, _ := cache.allocSlot()
-	cache.slotForSeq[2] = slot2
-	cache.refCount[slot2] = 1
-
-	initialFreeSlots := len(cache.freeSlots)
-
-	// Simulate overwriting seq 2's slot with slot1 (sharing)
-	// First free the old slot
-	cache.refCount[slot2]--
-	if cache.refCount[slot2] <= 0 {
-		cache.refCount[slot2] = 0
-		cache.freeSlot(slot2)
-	}
-	// Then share slot1
-	cache.slotForSeq[2] = slot1
-	cache.refCount[slot1]++
-
-	// Seq 2 should now share slot1
-	if cache.slotForSeq[2] != slot1 {
-		t.Errorf("expected seq 2 to share slot %d, got %d", slot1, cache.slotForSeq[2])
-	}
-
-	// Old slot2 should be freed
-	if len(cache.freeSlots) != initialFreeSlots+1 {
-		t.Errorf("expected %d free slots, got %d", initialFreeSlots+1, len(cache.freeSlots))
-	}
-}
-
-func TestHybridCache_BoundsChecking(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Test freeing invalid slot (should not panic)
-	cache.freeSlot(-1)
-	cache.freeSlot(100) // out of bounds
-
-	// freeSlot does bounds checking, so invalid slots should be ignored
-	if len(cache.freeSlots) != 4 {
-		t.Errorf("invalid slots should not affect free list, got %d slots", len(cache.freeSlots))
-	}
-}
-
-func TestHybridCache_MultipleSequences_RefCounting(t *testing.T) {
-	cache := createSlotOnlyCache(8)
-
-	// Allocate slot for seq 1
-	slot1, _ := cache.allocSlot()
-	cache.slotForSeq[1] = slot1
-	cache.refCount[slot1] = 1
-
-	// Fork to seq 2, 3, 4 (all share slot1)
-	for _, seq := range []int{2, 3, 4} {
-		cache.slotForSeq[seq] = slot1
-		cache.refCount[slot1]++
-	}
-
-	// Ref count should be 4
-	if cache.refCount[slot1] != 4 {
-		t.Errorf("expected refCount 4, got %d", cache.refCount[slot1])
-	}
-
-	// Remove seq 2, 3
-	for _, seq := range []int{2, 3} {
-		delete(cache.slotForSeq, seq)
-		cache.refCount[slot1]--
-	}
-
-	if cache.refCount[slot1] != 2 {
-		t.Errorf("expected refCount 2, got %d", cache.refCount[slot1])
-	}
-
-	// Slot should still be allocated (not in free list)
-	found := false
-	for _, s := range cache.freeSlots {
-		if s == slot1 {
-			found = true
-			break
-		}
-	}
-	if found {
-		t.Error("slot1 should not be in free list yet")
-	}
-
-	// Remove remaining sequences
-	for _, seq := range []int{1, 4} {
-		delete(cache.slotForSeq, seq)
-		cache.refCount[slot1]--
-	}
-
-	if cache.refCount[slot1] != 0 {
-		t.Errorf("expected refCount 0, got %d", cache.refCount[slot1])
-	}
-}
-
-func TestHybridCache_ChainedSharing(t *testing.T) {
-	cache := createSlotOnlyCache(8)
-
-	// Create seq 1
-	slot1, _ := cache.allocSlot()
-	cache.slotForSeq[1] = slot1
-	cache.refCount[slot1] = 1
-
-	// Share 1 -> 2
-	cache.slotForSeq[2] = slot1
-	cache.refCount[slot1]++
-
-	// Share 2 -> 3 (should still share slot1)
-	cache.slotForSeq[3] = cache.slotForSeq[2] // which is slot1
-	cache.refCount[slot1]++
-
-	// All should share slot1
-	if cache.slotForSeq[1] != slot1 || cache.slotForSeq[2] != slot1 || cache.slotForSeq[3] != slot1 {
-		t.Error("all sequences should share slot1")
-	}
-
-	if cache.refCount[slot1] != 3 {
-		t.Errorf("expected refCount 3, got %d", cache.refCount[slot1])
-	}
-}
-
-func TestHybridCache_CacheParameters(t *testing.T) {
-	cache := NewHybridCache(nil, 512, 5) // hiddenSize=512, dConv=5
-
-	if cache.hiddenSize != 512 {
-		t.Errorf("expected hiddenSize 512, got %d", cache.hiddenSize)
-	}
-	if cache.dConv != 5 {
-		t.Errorf("expected dConv 5, got %d", cache.dConv)
-	}
-}
-
-func TestHybridCache_NumSeqs(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Initially no sequences
-	if cache.numSeqs() != 0 {
-		t.Errorf("expected 0 seqs, got %d", cache.numSeqs())
-	}
-
-	// Manually set up current batch state
-	cache.curSeqs = []int{1, 2, 3}
-
-	if cache.numSeqs() != 3 {
-		t.Errorf("expected 3 seqs, got %d", cache.numSeqs())
-	}
-}
-
-func TestHybridCache_SeqTokens(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Initially 0
-	if cache.seqTokens() != 0 {
-		t.Errorf("expected 0 seqTokens, got %d", cache.seqTokens())
-	}
-
-	// Manually set up current batch state
-	cache.curSeqTokens = 16
-
-	if cache.seqTokens() != 16 {
-		t.Errorf("expected 16 seqTokens, got %d", cache.seqTokens())
-	}
-}
-
-// Test that Seqs returns a clone of curSeqs
-func TestHybridCache_Seqs_ReturnsClone(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	cache.curSeqs = []int{1, 2, 3}
-
-	seqs := cache.Seqs()
-
-	// Modify returned slice
-	seqs[0] = 999
-
-	// Original should be unchanged
-	if cache.curSeqs[0] != 1 {
-		t.Error("Seqs should return a clone, not the original slice")
-	}
-}
-
-func TestHybridCache_IsSupportedForBatch(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Initially not supported (no batch set up)
-	if cache.IsSupportedForBatch() {
-		t.Error("expected IsSupportedForBatch to be false initially")
-	}
-
-	// Set up a valid batch
-	cache.curSeqTokens = 1
-	cache.curSeqs = []int{1}
-
-	if !cache.IsSupportedForBatch() {
-		t.Error("expected IsSupportedForBatch to be true with valid batch")
-	}
-}
-
-func TestHybridCache_ZeroConvSlots_EmptyInputs(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// zeroConvSlots should handle empty slots without panicking
-	cache.zeroConvSlots(nil, nil)
-	cache.zeroConvSlots(nil, []int{})
-
-	// zeroConvSlots should handle empty convStates without panicking
-	cache.zeroConvSlots(nil, []int{0, 1, 2})
-}
-
-func TestHybridCache_SlotRecycling_TracksNewSlots(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Allocate slot for seq 1
-	slot1, _ := cache.allocSlot()
-	cache.slotForSeq[1] = slot1
-	cache.refCount[slot1] = 1
-
-	// Free the slot (simulating sequence removal)
-	cache.refCount[slot1]--
-	cache.freeSlot(slot1)
-	delete(cache.slotForSeq, 1)
-
-	// Verify slot is in free list
-	if len(cache.freeSlots) != 4 {
-		t.Errorf("expected 4 free slots after freeing, got %d", len(cache.freeSlots))
-	}
-
-	// Allocate for new seq 2 - should get recycled slot
-	slot2, _ := cache.allocSlot()
-	if slot2 != slot1 {
-		t.Errorf("expected recycled slot %d, got %d", slot1, slot2)
-	}
-
-	// This recycled slot would need zeroing in the real implementation
-	// The actual zeroing is tested via integration tests since it requires ML context
-}
-
-func TestHybridCache_NewSequence_GetsTrackedForZeroing(t *testing.T) {
-	cache := createSlotOnlyCache(4)
-
-	// Simulate the slot allocation flow from StartForward
-	// When a sequence doesn't have a slot, it gets allocated and tracked as "new"
-
-	newSlots := []int{}
-
-	// Seq 1 doesn't have a slot - allocate and track
-	seq := 1
-	if _, ok := cache.slotForSeq[seq]; !ok {
-		slot, err := cache.allocSlot()
-		if err != nil {
-			t.Fatalf("allocSlot failed: %v", err)
-		}
-		cache.slotForSeq[seq] = slot
-		cache.refCount[slot] = 1
-		newSlots = append(newSlots, slot)
-	}
-
-	// Verify newSlots contains the allocated slot
-	if len(newSlots) != 1 {
-		t.Errorf("expected 1 new slot, got %d", len(newSlots))
-	}
-
-	// Seq 1 already has a slot - should NOT be tracked as new
-	newSlots2 := []int{}
-	if _, ok := cache.slotForSeq[seq]; !ok {
-		slot, _ := cache.allocSlot()
-		cache.slotForSeq[seq] = slot
-		cache.refCount[slot] = 1
-		newSlots2 = append(newSlots2, slot)
-	}
-
-	// Verify no new slots for existing sequence
-	if len(newSlots2) != 0 {
-		t.Errorf("expected 0 new slots for existing sequence, got %d", len(newSlots2))
-	}
-}

model/models/lfm2/model.go

@@ -1,253 +0,0 @@
-package lfm2
-
-import (
-	"cmp"
-	"math"
-
-	"github.com/ollama/ollama/fs"
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/ml/nn"
-	"github.com/ollama/ollama/ml/nn/rope"
-	"github.com/ollama/ollama/model"
-	"github.com/ollama/ollama/model/input"
-)
-
-type Options struct {
-	hiddenSize       int
-	headDim, ropeDim int
-
-	eps, ropeBase, ropeScale float32
-
-	ropeType              string
-	originalContextLength int
-
-	// per-layer head counts (LFM2 alternates attention and recurrent layers)
-	numHeadsByLayer   []int
-	numKVHeadsByLayer []int
-}
-
-func (o Options) headDimValue() int {
-	// Head dim is shared across layers; fall back to first attention layer head count.
-	for _, h := range o.numHeadsByLayer {
-		if h > 0 {
-			return cmp.Or(o.headDim, o.hiddenSize/h)
-		}
-	}
-	return cmp.Or(o.headDim, o.hiddenSize)
-}
-
-func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor) ml.Tensor {
-	opts := []func(*rope.Options){rope.WithTypeNeoX()}
-	if o.ropeType == "yarn" {
-		attnFactor := float32(1.0 / (1.0 + 0.1*math.Log(float64(o.ropeScale))))
-		opts = append(opts,
-			rope.WithOriginalContextLength(o.originalContextLength),
-			rope.WithExtrapolationFactor(1.),
-			rope.WithAttentionFactor(attnFactor),
-		)
-	}
-
-	headCount := 1
-	for _, h := range o.numHeadsByLayer {
-		if h > 0 {
-			headCount = h
-			break
-		}
-	}
-	return nn.RoPE(ctx, states, positions, cmp.Or(o.ropeDim, o.headDim, o.hiddenSize/headCount), o.ropeBase, 1./o.ropeScale, opts...)
-}
-
-type Model struct {
-	model.Base
-	model.TextProcessor
-
-	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
-	Layers         []Layer       `gguf:"blk"`
-	OutputNorm     *nn.RMSNorm   `gguf:"output_norm,alt:token_embd_norm"`
-	Output         *nn.Linear    `gguf:"output,alt:token_embd"`
-
-	Options
-}
-
-func New(c fs.Config) (model.Model, error) {
-	if c.Uint("expert_count") > 0 {
-		return nil, model.ErrUnsupportedModel
-	}
-
-	if c.String("tokenizer.ggml.model") != "gpt2" {
-		return nil, model.ErrUnsupportedTokenizer
-	}
-
-	vocabulary := model.Vocabulary{
-		Values: c.Strings("tokenizer.ggml.tokens"),
-		Scores: c.Floats("tokenizer.ggml.scores"),
-		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")...,
-		),
-	}
-
-	var pretokenizers []string
-	switch c.String("tokenizer.ggml.pre") {
-	case "default":
-		// use default BPE pretokenizer
-	default:
-		// llama-bpe style (default for LFM2)
-		pretokenizers = []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+`,
-		}
-	}
-
-	m := Model{
-		TextProcessor: model.NewBytePairEncoding(&vocabulary, pretokenizers...),
-		Layers:        make([]Layer, c.Uint("block_count")),
-		Options: Options{
-			hiddenSize:            int(c.Uint("embedding_length")),
-			headDim:               int(c.Uint("attention.key_length")),
-			ropeDim:               int(c.Uint("rope.dimension_count")),
-			eps:                   c.Float("attention.layer_norm_rms_epsilon"),
-			ropeType:              c.String("rope.scaling.type"),
-			ropeBase:              c.Float("rope.freq_base"),
-			ropeScale:             c.Float("rope.scaling.factor", 1),
-			originalContextLength: int(c.Uint("rope.scaling.original_context_length")),
-		},
-	}
-
-	type headCounts interface {
-		HeadCount() []uint64
-		HeadCountKV() []uint64
-	}
-	hc, ok := c.(headCounts)
-	if !ok {
-		return nil, model.ErrUnsupportedModel
-	}
-
-	headCount := hc.HeadCount()
-	headCountKV := hc.HeadCountKV()
-
-	m.numHeadsByLayer = make([]int, len(m.Layers))
-	m.numKVHeadsByLayer = make([]int, len(m.Layers))
-	for i := range m.Layers {
-		m.numHeadsByLayer[i] = int(headCount[i])
-		m.numKVHeadsByLayer[i] = int(headCountKV[i])
-
-		if m.numKVHeadsByLayer[i] == 0 {
-			m.Layers[i].Operator = &ShortConv{}
-		} else {
-			m.Layers[i].Operator = &Attention{}
-		}
-	}
-
-	lCache := int(c.Uint("shortconv.l_cache"))
-	dConv := max(0, lCache-1)
-	m.Cache = NewHybridCache(m.Shift, m.hiddenSize, dConv)
-	return &m, nil
-}
-
-type Operator interface {
-	Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache *HybridCache, layer int, opts *Options) ml.Tensor
-}
-
-type Attention struct {
-	Query     *nn.Linear  `gguf:"attn_q"`
-	QueryNorm *nn.RMSNorm `gguf:"attn_q_norm"`
-	Key       *nn.Linear  `gguf:"attn_k"`
-	KeyNorm   *nn.RMSNorm `gguf:"attn_k_norm"`
-	Value     *nn.Linear  `gguf:"attn_v"`
-	Output    *nn.Linear  `gguf:"attn_output,alt:attn_out"`
-}
-
-func (sa *Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache *HybridCache, layer int, opts *Options) ml.Tensor {
-	batchSize := hiddenStates.Dim(1)
-	headDim := opts.headDimValue()
-	numHeads := opts.numHeadsByLayer[layer]
-	numKVHeads := opts.numKVHeadsByLayer[layer]
-
-	query := sa.Query.Forward(ctx, hiddenStates)
-	key := sa.Key.Forward(ctx, hiddenStates)
-	value := sa.Value.Forward(ctx, hiddenStates)
-
-	query = query.Reshape(ctx, headDim, numHeads, batchSize)
-	key = key.Reshape(ctx, headDim, numKVHeads, batchSize)
-	value = value.Reshape(ctx, headDim, numKVHeads, batchSize)
-
-	query = sa.QueryNorm.Forward(ctx, query, opts.eps)
-	key = sa.KeyNorm.Forward(ctx, key, opts.eps)
-
-	query = opts.applyRotaryPositionEmbeddings(ctx, query, positions)
-	key = opts.applyRotaryPositionEmbeddings(ctx, key, positions)
-
-	attention := nn.Attention(ctx, query, key, value, 1./math.Sqrt(float64(headDim)), cache)
-	attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), batchSize)
-	return sa.Output.Forward(ctx, attention)
-}
-
-type MLP struct {
-	Up   *nn.Linear `gguf:"ffn_up"`
-	Down *nn.Linear `gguf:"ffn_down"`
-	Gate *nn.Linear `gguf:"ffn_gate"`
-}
-
-func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *Options) ml.Tensor {
-	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx, mlp.Up.Forward(ctx, hiddenState))
-	return mlp.Down.Forward(ctx, hiddenState)
-}
-
-type Layer struct {
-	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
-	Operator      Operator
-	MLPNorm       *nn.RMSNorm `gguf:"ffn_norm"`
-	MLP           *MLP
-}
-
-func (l *Layer) Forward(ctx ml.Context, layer int, hiddenState, positions, outputs ml.Tensor, cache *HybridCache, opts *Options) ml.Tensor {
-	residual := hiddenState
-
-	hiddenState = l.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
-	hiddenState = l.Operator.Forward(ctx, hiddenState, positions, cache, layer, opts)
-
-	if outputs != nil {
-		hiddenState = hiddenState.Rows(ctx, outputs)
-		residual = residual.Rows(ctx, outputs)
-	}
-
-	hiddenState = hiddenState.Add(ctx, residual)
-	residual = hiddenState
-
-	hiddenState = l.MLPNorm.Forward(ctx, hiddenState, opts.eps)
-	hiddenState = l.MLP.Forward(ctx, hiddenState, opts)
-	return hiddenState.Add(ctx, residual)
-}
-
-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))
-
-	hiddenState := 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
-		}
-
-		hiddenState = layer.Forward(ctx, i, hiddenState, positions, outputs, m.Cache.(*HybridCache), &m.Options)
-	}
-
-	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
-	return m.Output.Forward(ctx, hiddenState), nil
-}
-
-func init() {
-	model.Register("lfm2", New)
-}

model/models/lfm2/shortconv.go

@@ -1,50 +0,0 @@
-package lfm2
-
-import (
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/ml/nn"
-)
-
-type shortConvKernel struct {
-	Weight ml.Tensor `gguf:"weight"`
-}
-
-// ShortConv implements the LFM2 short-convolution block (GGML_OP_SSM_CONV) with a recurrent
-// state stored in the HybridCache.
-type ShortConv struct {
-	Conv    *shortConvKernel `gguf:"shortconv.conv"`
-	InProj  *nn.Linear       `gguf:"shortconv.in_proj"`
-	OutProj *nn.Linear       `gguf:"shortconv.out_proj"`
-}
-
-func (sc *ShortConv) Forward(ctx ml.Context, hiddenStates ml.Tensor, _ ml.Tensor, cache *HybridCache, layer int, opts *Options) ml.Tensor {
-	nSeqs := cache.numSeqs()
-	seqTokens := cache.seqTokens()
-	hiddenSize := hiddenStates.Dim(0)
-	if nSeqs <= 0 || seqTokens <= 0 || hiddenStates.Dim(1) != nSeqs*seqTokens {
-		panic("lfm2: unsupported batch layout for shortconv")
-	}
-
-	bcx := sc.InProj.Forward(ctx, hiddenStates).Reshape(ctx, 3*hiddenSize, seqTokens, nSeqs)
-
-	elementSize := bcx.Stride(0)
-	b := bcx.View(ctx, 0*hiddenSize*elementSize, hiddenSize, bcx.Stride(1), seqTokens, bcx.Stride(2), nSeqs)
-	c := bcx.View(ctx, 1*hiddenSize*elementSize, hiddenSize, bcx.Stride(1), seqTokens, bcx.Stride(2), nSeqs)
-	x := bcx.View(ctx, 2*hiddenSize*elementSize, hiddenSize, bcx.Stride(1), seqTokens, bcx.Stride(2), nSeqs)
-
-	bx := b.Mul(ctx, x).Permute(ctx, 1, 0, 2, 3)
-
-	state, err := cache.ConvState(ctx, layer)
-	if err != nil {
-		panic("lfm2: failed to get conv state: " + err.Error())
-	}
-	sx := state.Concat(ctx, bx, 0)
-
-	convOut := sx.SSMConv(ctx, sc.Conv.Weight)
-	y := c.Mul(ctx, convOut)
-
-	dConv := sx.Dim(0) - seqTokens
-	cache.UpdateConvState(ctx, layer, sx.Slice(ctx, 0, sx.Dim(0)-dConv, sx.Dim(0), 1))
-
-	return sc.OutProj.Forward(ctx, y.Reshape(ctx, hiddenSize, seqTokens*nSeqs))
-}

model/models/models.go

@@ -7,9 +7,7 @@ import (
 	_ "github.com/ollama/ollama/model/models/gemma2"
 	_ "github.com/ollama/ollama/model/models/gemma3"
 	_ "github.com/ollama/ollama/model/models/gemma3n"
-	_ "github.com/ollama/ollama/model/models/glm4moelite"
 	_ "github.com/ollama/ollama/model/models/gptoss"
-	_ "github.com/ollama/ollama/model/models/lfm2"
 	_ "github.com/ollama/ollama/model/models/llama"
 	_ "github.com/ollama/ollama/model/models/llama4"
 	_ "github.com/ollama/ollama/model/models/mistral3"

model/parsers/glm46.go

@@ -1,410 +0,0 @@
-package parsers
-
-import (
-	"context"
-	"encoding/xml"
-	"fmt"
-	"log/slog"
-	"strings"
-	"unicode"
-
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/logutil"
-)
-
-type glm46ParserState int
-
-const (
-	glm46ParserState_LookingForThinkingOpen glm46ParserState = iota
-	glm46ParserState_ThinkingStartedEatingWhitespace
-	glm46ParserState_CollectingThinking
-	glm46ParserState_ThinkingDoneEatingWhitespace
-	glm46ParserState_CollectingContent
-	glm46ParserState_ToolStartedEatingWhitespace
-	glm46ParserState_CollectingToolContent
-)
-
-const (
-	glm46ThinkingOpenTag  = "<think>"
-	glm46ThinkingCloseTag = "</think>"
-	glm46ToolOpenTag      = "<tool_call>"
-	glm46ToolCloseTag     = "</tool_call>"
-)
-
-type GLM46Parser struct {
-	state  glm46ParserState
-	buffer strings.Builder
-	tools  []api.Tool
-}
-
-func (p *GLM46Parser) HasToolSupport() bool {
-	return true
-}
-
-func (p *GLM46Parser) HasThinkingSupport() bool {
-	return true
-}
-
-// func (p *GLM46Parser) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
-func (p *GLM46Parser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
-	p.tools = tools
-	return tools
-}
-
-type glm46Event interface {
-	isGLM46Event()
-}
-
-type glm46EventContent struct {
-	content string
-}
-
-func (glm46EventContent) isGLM46Event() {}
-
-type glm46EventRawToolCall struct {
-	raw string
-}
-
-func (glm46EventRawToolCall) isGLM46Event() {}
-
-type glm46EventThinkingContent struct {
-	content string
-}
-
-func (glm46EventThinkingContent) isGLM46Event() {}
-
-func (p *GLM46Parser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
-	p.buffer.WriteString(s)
-	events := p.parseEvents()
-
-	var toolCalls []api.ToolCall
-	var contentSb strings.Builder
-	var thinkingSb strings.Builder
-
-	for _, event := range events {
-		switch event := event.(type) {
-		case glm46EventRawToolCall:
-			toolCall, err := parseGLM46ToolCall(event, p.tools)
-			if err != nil {
-				slog.Warn("glm-4.6 tool call parsing failed", "error", err)
-				return "", "", nil, err
-			}
-			toolCalls = append(toolCalls, toolCall)
-		case glm46EventThinkingContent:
-			thinkingSb.WriteString(event.content)
-		case glm46EventContent:
-			// TODO(drifkin): if the same turn contains multiple interleaved content
-			// events, we naively append them together here.
-			contentSb.WriteString(event.content)
-		}
-	}
-
-	return contentSb.String(), thinkingSb.String(), toolCalls, nil
-}
-
-func (p *GLM46Parser) parseEvents() []glm46Event {
-	var all []glm46Event
-
-	keepLooping := true
-	for keepLooping {
-		var events []glm46Event
-		events, keepLooping = p.eat()
-		if len(events) > 0 {
-			all = append(all, events...)
-		}
-	}
-
-	if len(all) > 0 {
-		slog.Log(context.TODO(), logutil.LevelTrace, "glm-4.6 events parsed", "events", all, "state", p.state, "buffer", p.buffer.String())
-	}
-
-	return all
-}
-
-// eatLeadingWhitespaceAndTransitionTo consumes leading whitespace from the buffer
-// and transitions to the next state. Returns (nil, false) if only whitespace remains
-// in the buffer (needs more input), or (nil, true) if we successfully transitioned.
-func (p *GLM46Parser) eatLeadingWhitespaceAndTransitionTo(nextState glm46ParserState) ([]glm46Event, bool) {
-	trimmed := strings.TrimLeftFunc(p.buffer.String(), unicode.IsSpace)
-	p.buffer.Reset()
-	if trimmed == "" {
-		return nil, false // Still only whitespace, keep waiting for more input
-	}
-	p.state = nextState
-	p.buffer.WriteString(trimmed)
-	return nil, true // Successfully transitioned
-}
-
-// glm46SplitAtTag splits the buffer at the given tag, returns the content before (trimmed of trailing whitespace),
-// the content after (optionally trimmed of leading whitespace), and updates the buffer
-func glm46SplitAtTag(p *GLM46Parser, tag string, trimAfter bool) (string, string) {
-	split := strings.SplitN(p.buffer.String(), tag, 2)
-	before := split[0]
-	before = strings.TrimRightFunc(before, unicode.IsSpace)
-	after := split[1]
-	if trimAfter {
-		after = strings.TrimLeftFunc(after, unicode.IsSpace)
-	}
-	p.buffer.Reset()
-	p.buffer.WriteString(after)
-	return before, after
-}
-
-func (p *GLM46Parser) eat() ([]glm46Event, bool) {
-	var events []glm46Event
-
-	switch p.state {
-	case glm46ParserState_LookingForThinkingOpen:
-		trimmed := strings.TrimLeftFunc(p.buffer.String(), unicode.IsSpace)
-		if strings.HasPrefix(trimmed, glm46ThinkingOpenTag) {
-			// Found <think> opening tag
-			after := strings.TrimPrefix(trimmed, glm46ThinkingOpenTag)
-			after = strings.TrimLeftFunc(after, unicode.IsSpace)
-			p.buffer.Reset()
-			p.buffer.WriteString(after)
-			if after == "" {
-				p.state = glm46ParserState_ThinkingStartedEatingWhitespace
-			} else {
-				p.state = glm46ParserState_CollectingThinking
-			}
-			return events, true
-		} else if strings.HasPrefix(glm46ThinkingOpenTag, trimmed) {
-			// Partial opening tag seen, keep accumulating
-			return events, false
-		} else if trimmed == "" {
-			// Only whitespace, keep accumulating
-			return events, false
-		} else {
-			// No thinking tag found, skip to content collection
-			p.state = glm46ParserState_CollectingContent
-			// Don't trim - we want to keep the original content
-			return events, true
-		}
-
-	case glm46ParserState_ThinkingStartedEatingWhitespace:
-		return p.eatLeadingWhitespaceAndTransitionTo(glm46ParserState_CollectingThinking)
-
-	case glm46ParserState_CollectingThinking:
-		acc := p.buffer.String()
-		if strings.Contains(acc, glm46ThinkingCloseTag) {
-			thinking, remaining := glm46SplitAtTag(p, glm46ThinkingCloseTag, true)
-			if len(thinking) > 0 {
-				events = append(events, glm46EventThinkingContent{content: thinking})
-			}
-			if remaining == "" {
-				p.state = glm46ParserState_ThinkingDoneEatingWhitespace
-			} else {
-				p.state = glm46ParserState_CollectingContent
-			}
-			return events, true
-		} else if overlapLen := overlap(acc, glm46ThinkingCloseTag); overlapLen > 0 {
-			// Partial closing tag - withhold it along with any trailing whitespace before it
-			beforePartialTag := acc[:len(acc)-overlapLen]
-			trailingWhitespaceLen := trailingWhitespaceLen(beforePartialTag)
-			ambiguousStart := len(beforePartialTag) - trailingWhitespaceLen
-
-			unambiguous := acc[:ambiguousStart]
-			ambiguous := acc[ambiguousStart:]
-			p.buffer.Reset()
-			p.buffer.WriteString(ambiguous)
-			if len(unambiguous) > 0 {
-				events = append(events, glm46EventThinkingContent{content: unambiguous})
-			}
-			return events, false
-		} else {
-			// Pure thinking content - withhold trailing whitespace (might precede closing tag)
-			whitespaceLen := trailingWhitespaceLen(acc)
-			ambiguousStart := len(acc) - whitespaceLen
-
-			unambiguous := acc[:ambiguousStart]
-			ambiguous := acc[ambiguousStart:]
-			p.buffer.Reset()
-			p.buffer.WriteString(ambiguous)
-			if len(unambiguous) > 0 {
-				events = append(events, glm46EventThinkingContent{content: unambiguous})
-			}
-			return events, false
-		}
-
-	case glm46ParserState_ThinkingDoneEatingWhitespace:
-		return p.eatLeadingWhitespaceAndTransitionTo(glm46ParserState_CollectingContent)
-
-	case glm46ParserState_CollectingContent:
-		if strings.Contains(p.buffer.String(), glm46ToolOpenTag) {
-			before, after := glm46SplitAtTag(p, glm46ToolOpenTag, true)
-			if len(before) > 0 {
-				events = append(events, glm46EventContent{content: before})
-			}
-			if after == "" {
-				p.state = glm46ParserState_ToolStartedEatingWhitespace
-			} else {
-				p.state = glm46ParserState_CollectingToolContent
-			}
-			return events, true
-		} else if overlapLen := overlap(p.buffer.String(), glm46ToolOpenTag); overlapLen > 0 {
-			beforePartialTag := p.buffer.String()[:len(p.buffer.String())-overlapLen]
-			trailingWhitespaceLen := trailingWhitespaceLen(beforePartialTag)
-			ambiguousStart := len(beforePartialTag) - trailingWhitespaceLen
-
-			unambiguous := p.buffer.String()[:ambiguousStart]
-			ambiguous := p.buffer.String()[ambiguousStart:]
-			p.buffer.Reset()
-			p.buffer.WriteString(ambiguous)
-			if len(unambiguous) > 0 {
-				events = append(events, glm46EventContent{content: unambiguous})
-			}
-			return events, false
-		} else {
-			whitespaceLen := trailingWhitespaceLen(p.buffer.String())
-			ambiguousStart := len(p.buffer.String()) - whitespaceLen
-
-			unambiguous := p.buffer.String()[:ambiguousStart]
-			ambiguous := p.buffer.String()[ambiguousStart:]
-			p.buffer.Reset()
-			p.buffer.WriteString(ambiguous)
-			if len(unambiguous) > 0 {
-				events = append(events, glm46EventContent{content: unambiguous})
-			}
-			return events, false
-		}
-
-	case glm46ParserState_ToolStartedEatingWhitespace:
-		return p.eatLeadingWhitespaceAndTransitionTo(glm46ParserState_CollectingToolContent)
-
-	case glm46ParserState_CollectingToolContent:
-		acc := p.buffer.String()
-		if strings.Contains(acc, glm46ToolCloseTag) {
-			toolContent, _ := glm46SplitAtTag(p, glm46ToolCloseTag, true)
-			if len(toolContent) == 0 {
-				slog.Warn("glm46 tool call closing tag found but no content before it")
-			}
-			events = append(events, glm46EventRawToolCall{raw: toolContent})
-			p.state = glm46ParserState_CollectingContent
-			return events, true
-		} else {
-			// Keep accumulating - tool calls are not streamed
-			// We just wait for the closing tag
-			return events, false
-		}
-
-	default:
-		panic("unreachable")
-	}
-}
-
-// GLMToolCallXML represents the structure of a GLM-4.6 tool call for XML parsing
-type GLMToolCallXML struct {
-	XMLName xml.Name `xml:"tool_call"`
-	Content string   `xml:",chardata"` // Function name (text nodes between tags)
-	Keys    []string `xml:"arg_key"`   // All arg_key elements in document order
-	Values  []string `xml:"arg_value"` // All arg_value elements in document order
-}
-
-// escapeGLM46Content escapes XML entities in text content while preserving arg_key/arg_value tags
-func escapeGLM46Content(s string) string {
-	var result strings.Builder
-	inTag := false
-
-	for i := range len(s) {
-		ch := s[i]
-
-		if ch == '<' {
-			// Check if this is a known tag
-			if strings.HasPrefix(s[i:], "<arg_key>") ||
-				strings.HasPrefix(s[i:], "</arg_key>") ||
-				strings.HasPrefix(s[i:], "<arg_value>") ||
-				strings.HasPrefix(s[i:], "</arg_value>") {
-				inTag = true
-			}
-		}
-
-		if inTag {
-			result.WriteByte(ch)
-			if ch == '>' {
-				inTag = false
-			}
-		} else {
-			// Escape special characters in text content
-			switch ch {
-			case '&':
-				result.WriteString("&amp;")
-			case '<':
-				result.WriteString("&lt;")
-			case '>':
-				result.WriteString("&gt;")
-			default:
-				result.WriteByte(ch)
-			}
-		}
-	}
-
-	return result.String()
-}
-
-func parseGLM46ToolCall(raw glm46EventRawToolCall, tools []api.Tool) (api.ToolCall, error) {
-	// Escape any unescaped entities in text content
-	// We need to escape text between tags, but not the tags themselves
-	escaped := escapeGLM46Content(raw.raw)
-
-	// Wrap the content in a root element to make it valid XML
-	xmlString := "<tool_call>" + escaped + "</tool_call>"
-
-	// Parse XML into struct
-	var parsed GLMToolCallXML
-	if err := xml.Unmarshal([]byte(xmlString), &parsed); err != nil {
-		return api.ToolCall{}, fmt.Errorf("failed to parse XML: %w", err)
-	}
-
-	// Extract and trim function name
-	functionName := strings.TrimSpace(parsed.Content)
-	if functionName == "" {
-		return api.ToolCall{}, fmt.Errorf("empty function name")
-	}
-
-	// Verify keys and values are paired correctly
-	if len(parsed.Keys) != len(parsed.Values) {
-		return api.ToolCall{}, fmt.Errorf("mismatched arg_key and arg_value counts: %d keys, %d values", len(parsed.Keys), len(parsed.Values))
-	}
-
-	// Find the matching tool to get parameter types
-	var matchedTool *api.Tool
-	for i := range tools {
-		if tools[i].Function.Name == functionName {
-			matchedTool = &tools[i]
-			break
-		}
-	}
-
-	// Build arguments map by pairing keys and values
-	toolCall := api.ToolCall{
-		Function: api.ToolCallFunction{
-			Name:      functionName,
-			Arguments: api.NewToolCallFunctionArguments(),
-		},
-	}
-
-	for i := range parsed.Keys {
-		key := strings.TrimSpace(parsed.Keys[i])
-		value := parsed.Values[i] // Don't trim here - parseValue handles it
-
-		// Look up parameter type
-		var paramType api.PropertyType
-		if matchedTool != nil && matchedTool.Function.Parameters.Properties != nil {
-			if prop, ok := matchedTool.Function.Parameters.Properties.Get(key); ok {
-				// Handle anyOf by collecting all types from the union
-				if len(prop.AnyOf) > 0 {
-					for _, anyOfProp := range prop.AnyOf {
-						paramType = append(paramType, anyOfProp.Type...)
-					}
-				} else {
-					paramType = prop.Type
-				}
-			}
-		}
-
-		// Parse value with type coercion
-		toolCall.Function.Arguments.Set(key, parseValue(value, paramType))
-	}
-
-	return toolCall, nil
-}

model/parsers/glm46_test.go

@@ -1,862 +0,0 @@
-package parsers
-
-import (
-	"encoding/xml"
-	"reflect"
-	"testing"
-
-	"github.com/ollama/ollama/api"
-)
-
-func TestGLM46ParserStreaming(t *testing.T) {
-	type step struct {
-		input      string
-		wantEvents []glm46Event
-	}
-
-	cases := []struct {
-		desc  string
-		steps []step
-		only  bool
-	}{
-		{
-			desc: "leading whitespace before think tag",
-			steps: []step{
-				{
-					input:      "   \n\t  ",
-					wantEvents: []glm46Event{},
-				},
-				{
-					input:      "<think>thinking</think>",
-					wantEvents: []glm46Event{glm46EventThinkingContent{content: "thinking"}},
-				},
-			},
-		},
-		{
-			desc: "think tag with whitespace inside",
-			steps: []step{
-				{
-					input: "<think>  \n  thinking content  \n  </think>regular content",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "thinking content"},
-						glm46EventContent{content: "regular content"},
-					},
-				},
-			},
-		},
-		{
-			desc: "tool call with leading whitespace after opening tag",
-			steps: []step{
-				{
-					input: "<think></think><tool_call>  \n  test  \n  </tool_call>",
-					wantEvents: []glm46Event{
-						glm46EventRawToolCall{raw: "test"},
-					},
-				},
-			},
-		},
-		{
-			desc: "simple thinking then content",
-			steps: []step{
-				{
-					input: "<think>I am thinking</think>Now I respond",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "I am thinking"},
-						glm46EventContent{content: "Now I respond"},
-					},
-				},
-			},
-		},
-		{
-			desc: "streamed thinking content",
-			steps: []step{
-				{
-					input:      "<think>hello",
-					wantEvents: []glm46Event{glm46EventThinkingContent{content: "hello"}},
-				},
-				{
-					input:      " world",
-					wantEvents: []glm46Event{glm46EventThinkingContent{content: " world"}},
-				},
-				{
-					input: "</think>content",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "content"},
-					},
-				},
-			},
-		},
-		{
-			desc: "content before tool call",
-			steps: []step{
-				{
-					input: "<think>Let me call a tool</think>here is text<tool_call>",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "Let me call a tool"},
-						glm46EventContent{content: "here is text"},
-					},
-				},
-				{
-					input: "function_name\n<arg_key>param</arg_key>\n<arg_value>value</arg_value>\n</tool_call>",
-					wantEvents: []glm46Event{
-						glm46EventRawToolCall{raw: "function_name\n<arg_key>param</arg_key>\n<arg_value>value</arg_value>"},
-					},
-				},
-			},
-		},
-		{
-			desc: "tool call with content after",
-			steps: []step{
-				{
-					input: "<think>thinking</think><tool_call>test</tool_call>after tool",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "thinking"},
-						glm46EventRawToolCall{raw: "test"},
-						glm46EventContent{content: "after tool"},
-					},
-				},
-			},
-		},
-		{
-			desc: "trailing whitespace between content and tool call is trimmed",
-			steps: []step{
-				{
-					input: "<think>thinking</think>content\n  \t  <tool_call>test</tool_call>",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "thinking"},
-						glm46EventContent{content: "content"},
-						glm46EventRawToolCall{raw: "test"},
-					},
-				},
-			},
-		},
-		{
-			desc: "trailing whitespace between tool call and content is trimmed",
-			steps: []step{
-				{
-					input: "<think>think</think><tool_call>test</tool_call>\n\t  after",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "think"},
-						glm46EventRawToolCall{raw: "test"},
-						glm46EventContent{content: "after"},
-					},
-				},
-			},
-		},
-		{
-			desc: "split thinking close tag",
-			steps: []step{
-				{
-					input:      "<think>thinking content</th",
-					wantEvents: []glm46Event{glm46EventThinkingContent{content: "thinking content"}},
-				},
-				{
-					input: "ink>after",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "after"},
-					},
-				},
-			},
-		},
-		{
-			desc: "split thinking open tag",
-			steps: []step{
-				{
-					input:      "  <thi",
-					wantEvents: []glm46Event{},
-				},
-				{
-					input:      "nk>content</think>",
-					wantEvents: []glm46Event{glm46EventThinkingContent{content: "content"}},
-				},
-			},
-		},
-		{
-			desc: "split tool open tag",
-			steps: []step{
-				{
-					input:      "<think>think</think>content<tool",
-					wantEvents: []glm46Event{glm46EventThinkingContent{content: "think"}, glm46EventContent{content: "content"}},
-				},
-				{
-					input:      "_call>inside",
-					wantEvents: []glm46Event{},
-				},
-				{
-					input: "</tool_call>",
-					wantEvents: []glm46Event{
-						glm46EventRawToolCall{raw: "inside"},
-					},
-				},
-			},
-		},
-		{
-			desc: "partial thinking close tag fakeout",
-			steps: []step{
-				{
-					input:      "<think>content</th",
-					wantEvents: []glm46Event{glm46EventThinkingContent{content: "content"}},
-				},
-				{
-					input:      "ought more",
-					wantEvents: []glm46Event{glm46EventThinkingContent{content: "</thought more"}},
-				},
-			},
-		},
-		{
-			desc: "partial thinking open tag fakeout",
-			steps: []step{
-				{
-					input:      "  <thi",
-					wantEvents: []glm46Event{},
-				},
-				{
-					input: "nking is fun",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "  <thinking is fun"},
-					},
-				},
-			},
-		},
-		{
-			desc: "partial tool open tag fakeout",
-			steps: []step{
-				{
-					input: "<think></think>content\n<tool",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "content"},
-					},
-				},
-				{
-					input: " fakeout",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "\n<tool fakeout"},
-					},
-				},
-			},
-		},
-		{
-			desc: "partial tool close tag fakeout",
-			steps: []step{
-				{
-					input:      "<think></think><tool_call>content</tool",
-					wantEvents: []glm46Event{},
-				},
-				{
-					input:      " fakeout",
-					wantEvents: []glm46Event{},
-				},
-				{
-					input: "</tool_call>",
-					wantEvents: []glm46Event{
-						glm46EventRawToolCall{raw: "content</tool fakeout"},
-					},
-				},
-			},
-		},
-		{
-			desc: "empty thinking tag",
-			steps: []step{
-				{
-					input: "<think></think>content here",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "content here"},
-					},
-				},
-			},
-		},
-		{
-			desc: "multiple tool calls in sequence",
-			steps: []step{
-				{
-					input: "<think>think</think><tool_call>first</tool_call>between<tool_call>second</tool_call>end",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "think"},
-						glm46EventRawToolCall{raw: "first"},
-						glm46EventContent{content: "between"},
-						glm46EventRawToolCall{raw: "second"},
-						glm46EventContent{content: "end"},
-					},
-				},
-			},
-		},
-		{
-			desc: "no thinking tag - direct to content",
-			steps: []step{
-				{
-					input: "just content here",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "just content here"},
-					},
-				},
-			},
-		},
-		{
-			desc: "no thinking tag - skip to content then tool call",
-			steps: []step{
-				{
-					input: "Here's the answer:<tool_call>test</tool_call>done",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "Here's the answer:"},
-						glm46EventRawToolCall{raw: "test"},
-						glm46EventContent{content: "done"},
-					},
-				},
-			},
-		},
-		{
-			desc: "no thinking tag - whitespace preserved when no tags",
-			steps: []step{
-				{
-					input: "  \n  content with leading whitespace",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "  \n  content with leading whitespace"},
-					},
-				},
-			},
-		},
-		{
-			desc: "whitespace after think close tag gets eaten",
-			steps: []step{
-				{
-					input: "<think>thinking</think>  \n\t  content",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "thinking"},
-						glm46EventContent{content: "content"},
-					},
-				},
-			},
-		},
-		{
-			desc: "whitespace after tool_call close tag gets eaten",
-			steps: []step{
-				{
-					input: "<think></think><tool_call>test</tool_call>  \n\t  content",
-					wantEvents: []glm46Event{
-						glm46EventRawToolCall{raw: "test"},
-						glm46EventContent{content: "content"},
-					},
-				},
-			},
-		},
-		{
-			desc: "thinking content withholds trailing whitespace (single chunk)",
-			steps: []step{
-				{
-					input: "<think>thinking content   ",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "thinking content"},
-					},
-				},
-				{
-					input: "</think>after",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "after"},
-					},
-				},
-			},
-		},
-		{
-			desc: "thinking content withholds trailing whitespace with newlines",
-			steps: []step{
-				{
-					input: "<think>thinking\n\n  ",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "thinking"},
-					},
-				},
-				{
-					input: "</think>content",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "content"},
-					},
-				},
-			},
-		},
-		{
-			desc: "thinking content trailing whitespace emitted when more content arrives",
-			steps: []step{
-				{
-					input: "<think>thinking   ",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "thinking"},
-					},
-				},
-				{
-					input: "more thinking",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "   more thinking"},
-					},
-				},
-				{
-					input:      "</think>",
-					wantEvents: []glm46Event{},
-				},
-			},
-		},
-		{
-			desc: "thinking content withholds trailing whitespace before partial close tag",
-			steps: []step{
-				{
-					input: "<think>thinking   </th",
-					wantEvents: []glm46Event{
-						glm46EventThinkingContent{content: "thinking"},
-					},
-				},
-				{
-					input: "ink>content",
-					wantEvents: []glm46Event{
-						glm46EventContent{content: "content"},
-					},
-				},
-			},
-		},
-	}
-
-	anyOnlies := false
-	for _, tc := range cases {
-		if tc.only {
-			anyOnlies = true
-		}
-	}
-
-	for _, tc := range cases {
-		if anyOnlies && !tc.only {
-			continue
-		}
-
-		t.Run(tc.desc, func(t *testing.T) {
-			parser := GLM46Parser{}
-
-			for i, step := range tc.steps {
-				parser.buffer.WriteString(step.input)
-				gotEvents := parser.parseEvents()
-
-				if len(gotEvents) == 0 && len(step.wantEvents) == 0 {
-					// avoid deep equal on empty vs. nil slices
-					continue
-				}
-
-				if !reflect.DeepEqual(gotEvents, step.wantEvents) {
-					t.Errorf("step %d: input %q: got events %#v, want %#v", i, step.input, gotEvents, step.wantEvents)
-				}
-			}
-		})
-	}
-}
-
-// TestGLMToolCallXMLOrderPreservation verifies that xml.Unmarshal preserves
-// document order when collecting multiple elements with the same tag name into slices.
-// This is a critical assumption for the GLM-4.6 parser's struct-based approach.
-func TestGLMToolCallXMLOrderPreservation(t *testing.T) {
-	testCases := []struct {
-		name       string
-		xml        string
-		wantKeys   []string
-		wantValues []string
-	}{
-		{
-			name: "alternating keys and values",
-			xml: `<tool_call>
-function_name
-<arg_key>first</arg_key>
-<arg_value>A</arg_value>
-<arg_key>second</arg_key>
-<arg_value>B</arg_value>
-<arg_key>third</arg_key>
-<arg_value>C</arg_value>
-</tool_call>`,
-			wantKeys:   []string{"first", "second", "third"},
-			wantValues: []string{"A", "B", "C"},
-		},
-		{
-			name: "all keys then all values",
-			xml: `<tool_call>
-function_name
-<arg_key>key1</arg_key>
-<arg_key>key2</arg_key>
-<arg_key>key3</arg_key>
-<arg_value>val1</arg_value>
-<arg_value>val2</arg_value>
-<arg_value>val3</arg_value>
-</tool_call>`,
-			wantKeys:   []string{"key1", "key2", "key3"},
-			wantValues: []string{"val1", "val2", "val3"},
-		},
-		{
-			name: "mixed grouping",
-			xml: `<tool_call>
-function_name
-<arg_key>a</arg_key>
-<arg_value>1</arg_value>
-<arg_key>b</arg_key>
-<arg_key>c</arg_key>
-<arg_value>2</arg_value>
-<arg_value>3</arg_value>
-</tool_call>`,
-			wantKeys:   []string{"a", "b", "c"},
-			wantValues: []string{"1", "2", "3"},
-		},
-		{
-			name: "reverse order - all values then all keys",
-			xml: `<tool_call>
-function_name
-<arg_value>X</arg_value>
-<arg_value>Y</arg_value>
-<arg_value>Z</arg_value>
-<arg_key>x</arg_key>
-<arg_key>y</arg_key>
-<arg_key>z</arg_key>
-</tool_call>`,
-			wantKeys:   []string{"x", "y", "z"},
-			wantValues: []string{"X", "Y", "Z"},
-		},
-	}
-
-	for _, tc := range testCases {
-		t.Run(tc.name, func(t *testing.T) {
-			var parsed GLMToolCallXML
-			err := xml.Unmarshal([]byte(tc.xml), &parsed)
-			if err != nil {
-				t.Fatalf("failed to unmarshal XML: %v", err)
-			}
-
-			if !reflect.DeepEqual(parsed.Keys, tc.wantKeys) {
-				t.Errorf("Keys order mismatch:\ngot:  %v\nwant: %v", parsed.Keys, tc.wantKeys)
-			}
-
-			if !reflect.DeepEqual(parsed.Values, tc.wantValues) {
-				t.Errorf("Values order mismatch:\ngot:  %v\nwant: %v", parsed.Values, tc.wantValues)
-			}
-		})
-	}
-}
-
-func TestGLM46ToolCallParsing(t *testing.T) {
-	type testCase struct {
-		name         string
-		rawToolCall  string
-		tools        []api.Tool
-		wantToolCall api.ToolCall
-	}
-
-	cases := []testCase{
-		{
-			name:  "simple tool call",
-			tools: []api.Tool{},
-			rawToolCall: `get-current-weather
-<arg_key>location</arg_key>
-<arg_value>New York, NY</arg_value>
-<arg_key>unit</arg_key>
-<arg_value>celsius</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "get-current-weather",
-					Arguments: args(`{"location": "New York, NY", "unit": "celsius"}`),
-				},
-			},
-		},
-		{
-			name: "tool call with typed parameters",
-			tools: []api.Tool{
-				tool("calculate", map[string]api.ToolProperty{
-					"x":       {Type: api.PropertyType{"number"}},
-					"y":       {Type: api.PropertyType{"integer"}},
-					"enabled": {Type: api.PropertyType{"boolean"}},
-					"items":   {Type: api.PropertyType{"array"}},
-				}),
-			},
-			rawToolCall: `calculate
-<arg_key>x</arg_key>
-<arg_value>3.14</arg_value>
-<arg_key>y</arg_key>
-<arg_value>42</arg_value>
-<arg_key>enabled</arg_key>
-<arg_value>true</arg_value>
-<arg_key>items</arg_key>
-<arg_value>["a", "b", "c"]</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "calculate",
-					Arguments: args(`{"enabled": true, "items": ["a", "b", "c"], "x": 3.14, "y": 42}`),
-				},
-			},
-		},
-		{
-			name:  "function name with whitespace",
-			tools: []api.Tool{},
-			rawToolCall: `  get-weather
-<arg_key>city</arg_key>
-<arg_value>Paris</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "get-weather",
-					Arguments: args(`{"city": "Paris"}`),
-				},
-			},
-		},
-		{
-			name:  "values with special characters",
-			tools: []api.Tool{},
-			rawToolCall: `execute-command
-<arg_key>command</arg_key>
-<arg_value>ls && echo "done"</arg_value>
-<arg_key>message</arg_key>
-<arg_value>a < b and c > d</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "execute-command",
-					Arguments: args(`{"command": "ls && echo \"done\"", "message": "a < b and c > d"}`),
-				},
-			},
-		},
-		{
-			name:  "unicode in function names and values",
-			tools: []api.Tool{},
-			rawToolCall: `获取天气
-<arg_key>城市</arg_key>
-<arg_value>北京</arg_value>
-<arg_key>message</arg_key>
-<arg_value>Hello! 你好! 🌟</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "获取天气",
-					Arguments: args(`{"message": "Hello! 你好! 🌟", "城市": "北京"}`),
-				},
-			},
-		},
-		{
-			name:  "empty value",
-			tools: []api.Tool{},
-			rawToolCall: `test-function
-<arg_key>param1</arg_key>
-<arg_value></arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "test-function",
-					Arguments: args(`{"param1": ""}`),
-				},
-			},
-		},
-		{
-			name:  "special chars in arg_key names",
-			tools: []api.Tool{},
-			rawToolCall: `test-function
-<arg_key>param<1></arg_key>
-<arg_value>value1</arg_value>
-<arg_key>a&b</arg_key>
-<arg_value>value2</arg_value>
-<arg_key>x>y</arg_key>
-<arg_value>value3</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "test-function",
-					Arguments: args(`{"a&b": "value2", "param<1>": "value1", "x>y": "value3"}`),
-				},
-			},
-		},
-		{
-			name:  "multiple consecutive ampersands",
-			tools: []api.Tool{},
-			rawToolCall: `test-function
-<arg_key>param</arg_key>
-<arg_value>test &&&& more</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "test-function",
-					Arguments: args(`{"param": "test &&&& more"}`),
-				},
-			},
-		},
-		{
-			name:  "mixed special chars together",
-			tools: []api.Tool{},
-			rawToolCall: `test-function
-<arg_key>param</arg_key>
-<arg_value><>&<>&</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "test-function",
-					Arguments: args(`{"param": "<>&<>&"}`),
-				},
-			},
-		},
-		{
-			name:  "newlines and tabs in parameter values",
-			tools: []api.Tool{},
-			rawToolCall: `test-function
-<arg_key>multiline</arg_key>
-<arg_value>line1
-	indented line2
-line3</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "test-function",
-					Arguments: args(`{"multiline": "line1\n\tindented line2\nline3"}`),
-				},
-			},
-		},
-		{
-			name:  "single and double quotes in values",
-			tools: []api.Tool{},
-			rawToolCall: `test-function
-<arg_key>quotes</arg_key>
-<arg_value>She said "Hello's there!"</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "test-function",
-					Arguments: args(`{"quotes": "She said \"Hello's there!\""}`),
-				},
-			},
-		},
-		{
-			name:  "CDATA-like content that should be treated as text",
-			tools: []api.Tool{},
-			rawToolCall: `test-function
-<arg_key>cdata</arg_key>
-<arg_value><![CDATA[not actual cdata]]></arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "test-function",
-					Arguments: args(`{"cdata": "<![CDATA[not actual cdata]]>"}`),
-				},
-			},
-		},
-		{
-			name:  "all special XML entities",
-			tools: []api.Tool{},
-			rawToolCall: `test-function
-<arg_key>entities</arg_key>
-<arg_value>&lt;&gt;&amp;&apos;&quot;</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "test-function",
-					Arguments: args(`{"entities": "&lt;&gt;&amp;&apos;&quot;"}`),
-				},
-			},
-		},
-		{
-			name:  "order preservation with multiple parameters",
-			tools: []api.Tool{},
-			rawToolCall: `test-function
-<arg_key>first</arg_key>
-<arg_value>value1</arg_value>
-<arg_key>second</arg_key>
-<arg_value>value2</arg_value>
-<arg_key>third</arg_key>
-<arg_value>value3</arg_value>
-<arg_key>fourth</arg_key>
-<arg_value>value4</arg_value>
-<arg_key>fifth</arg_key>
-<arg_value>value5</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "test-function",
-					Arguments: args(`{"fifth": "value5", "first": "value1", "fourth": "value4", "second": "value2", "third": "value3"}`),
-				},
-			},
-		},
-		{
-			name:  "order preservation with identical key names but different positions",
-			tools: []api.Tool{},
-			rawToolCall: `test-function
-<arg_key>param</arg_key>
-<arg_value>first occurrence</arg_value>
-<arg_key>other</arg_key>
-<arg_value>middle</arg_value>
-<arg_key>param</arg_key>
-<arg_value>second occurrence</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name: "test-function",
-					// Later occurrence should overwrite earlier one
-					Arguments: args(`{"other": "middle", "param": "second occurrence"}`),
-				},
-			},
-		},
-		{
-			name: "array with mixed types",
-			tools: []api.Tool{
-				tool("process", map[string]api.ToolProperty{
-					"items": {Type: api.PropertyType{"array"}},
-				}),
-			},
-			rawToolCall: `process
-<arg_key>items</arg_key>
-<arg_value>[1, "hello", true, null]</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "process",
-					Arguments: args(`{"items": [1, "hello", true, null]}`),
-				},
-			},
-		},
-		{
-			name: "empty array",
-			tools: []api.Tool{
-				tool("test", map[string]api.ToolProperty{
-					"tags": {Type: api.PropertyType{"array"}},
-				}),
-			},
-			rawToolCall: `test
-<arg_key>tags</arg_key>
-<arg_value>[]</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "test",
-					Arguments: args(`{"tags": []}`),
-				},
-			},
-		},
-		{
-			name: "anyOf array or string - with array of objects",
-			tools: []api.Tool{
-				tool("TodoWrite", map[string]api.ToolProperty{
-					"todos": {AnyOf: []api.ToolProperty{{Type: api.PropertyType{"array"}}, {Type: api.PropertyType{"string"}}}},
-				}),
-			},
-			// <tool_call>TodoWrite
-			// <arg_key>todos</arg_key>
-			// <arg_value>[{"content": "Set up HTML file and basic structure", "id": "1", "priority": "high", "status": "pending"}, {"content": "Create 3D scene with Three.js", "id": "2", "priority": "high", "status": "pending"}, {"content": "Implement terrain generation with blocks", "id": "3", "priority": "high", "status": "pending"}, {"content": "Add player controls (movement, camera)", "id": "4", "priority": "high", "status": "pending"}, {"content": "Implement block placement/destruction", "id": "5", "priority": "medium", "status": "pending"}, {"content": "Add lighting and textures", "id": "6", "priority": "medium", "status": "pending"}, {"content": "Test and optimize performance", "id": "7", "priority": "low", "status": "pending"}]</arg_value>
-			// </tool_call>
-			rawToolCall: `TodoWrite
-<arg_key>todos</arg_key>
-<arg_value>[{"content": "task 1", "status": "pending", "priority": "high", "id": "1"}, {"content": "task 2", "status": "completed", "priority": "low", "id": "2"}]</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "TodoWrite",
-					Arguments: args(`{"todos": [{"content": "task 1", "id": "1", "priority": "high", "status": "pending"}, {"content": "task 2", "id": "2", "priority": "low", "status": "completed"}]}`),
-				},
-			},
-		},
-		{
-			name: "anyOf array or string - with plain string",
-			tools: []api.Tool{
-				tool("TodoWrite", map[string]api.ToolProperty{
-					"todos": {Type: api.PropertyType{"array", "string"}},
-				}),
-			},
-			rawToolCall: `TodoWrite
-<arg_key>todos</arg_key>
-<arg_value>Error: could not load todos</arg_value>`,
-			wantToolCall: api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      "TodoWrite",
-					Arguments: args(`{"todos": "Error: could not load todos"}`),
-				},
-			},
-		},
-	}
-
-	for i, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			gotToolCall, err := parseGLM46ToolCall(glm46EventRawToolCall{raw: tc.rawToolCall}, tc.tools)
-			if err != nil {
-				t.Errorf("case %d (%s): %v", i, tc.name, err)
-			}
-			if !toolCallEqual(gotToolCall, tc.wantToolCall) {
-				t.Errorf("case %d (%s): got tool call %#v, want %#v", i, tc.name, gotToolCall, tc.wantToolCall)
-			}
-		})
-	}
-}

model/parsers/glm47.go

@@ -1,20 +0,0 @@
-package parsers
-
-import "github.com/ollama/ollama/api"
-
-// GLM47Parser extends GLM46Parser with thinking-aware initialization.
-// GLM-4.7's prompt ends with <think> when thinking is enabled, so the parser
-// must start in CollectingThinking state (the model outputs thinking content directly).
-type GLM47Parser struct {
-	GLM46Parser
-}
-
-func (p *GLM47Parser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
-	p.tools = tools
-	// When thinking is enabled (nil or true), the prompt ends with <think>,
-	// so model output starts directly with thinking content (no opening tag).
-	if thinkValue == nil || thinkValue.Bool() {
-		p.state = glm46ParserState_CollectingThinking
-	}
-	return tools
-}

model/parsers/glm47_test.go

@@ -1,99 +0,0 @@
-package parsers
-
-import (
-	"reflect"
-	"testing"
-
-	"github.com/ollama/ollama/api"
-)
-
-func TestGLM47ParserAdd(t *testing.T) {
-	parser := GLM47Parser{}
-	parser.Init([]api.Tool{
-		tool("calculate", map[string]api.ToolProperty{
-			"count":   {Type: api.PropertyType{"integer"}},
-			"enabled": {Type: api.PropertyType{"boolean"}},
-		}),
-	}, nil, nil)
-
-	// When thinking is enabled (thinkValue nil), the prompt ends with <think>,
-	// so the model output does NOT include the opening <think> tag.
-	content, thinking, calls, err := parser.Add("plan</think>Answer<tool_call>calculate<arg_key>count</arg_key><arg_value>3</arg_value><arg_key>enabled</arg_key><arg_value>true</arg_value></tool_call>", true)
-	if err != nil {
-		t.Fatalf("parse failed: %v", err)
-	}
-	if thinking != "plan" {
-		t.Fatalf("expected thinking 'plan', got %q", thinking)
-	}
-	if content != "Answer" {
-		t.Fatalf("expected content 'Answer', got %q", content)
-	}
-	if len(calls) != 1 {
-		t.Fatalf("expected 1 tool call, got %d", len(calls))
-	}
-	expectedArgs := args(`{"count": 3, "enabled": true}`)
-	if !toolCallEqual(api.ToolCall{Function: api.ToolCallFunction{Arguments: calls[0].Function.Arguments}}, api.ToolCall{Function: api.ToolCallFunction{Arguments: expectedArgs}}) {
-		t.Fatalf("expected args %#v, got %#v", expectedArgs.ToMap(), calls[0].Function.Arguments.ToMap())
-	}
-}
-
-func TestGLM47ParserNoThinkingContent(t *testing.T) {
-	parser := GLM47Parser{}
-	parser.Init(nil, nil, nil)
-
-	// When thinking is enabled but model has no thinking to output,
-	// it should output </think> immediately followed by content.
-	content, thinking, calls, err := parser.Add("</think>Plain answer", true)
-	if err != nil {
-		t.Fatalf("parse failed: %v", err)
-	}
-	if thinking != "" {
-		t.Fatalf("expected empty thinking, got %q", thinking)
-	}
-	if content != "Plain answer" {
-		t.Fatalf("expected content 'Plain answer', got %q", content)
-	}
-	if len(calls) != 0 {
-		t.Fatalf("expected no tool calls, got %d", len(calls))
-	}
-}
-
-func TestGLM47ParserThinkingDisabled(t *testing.T) {
-	parser := GLM47Parser{}
-	// When thinking is disabled, parser stays in LookingForThinkingOpen state
-	parser.Init(nil, nil, &api.ThinkValue{Value: false})
-
-	// Model outputs plain content (prompt ended with </think>)
-	content, thinking, calls, err := parser.Add("Plain answer", true)
-	if err != nil {
-		t.Fatalf("parse failed: %v", err)
-	}
-	if thinking != "" {
-		t.Fatalf("expected empty thinking, got %q", thinking)
-	}
-	if content != "Plain answer" {
-		t.Fatalf("expected content 'Plain answer', got %q", content)
-	}
-	if len(calls) != 0 {
-		t.Fatalf("expected no tool calls, got %d", len(calls))
-	}
-}
-
-func TestGLM47ParserToolCallEscaping(t *testing.T) {
-	toolCall, err := parseGLM46ToolCall(glm46EventRawToolCall{raw: `exec
-<arg_key>expr</arg_key>
-<arg_value>a < b && c > d</arg_value>`}, nil)
-	if err != nil {
-		t.Fatalf("parse failed: %v", err)
-	}
-
-	expected := api.ToolCall{
-		Function: api.ToolCallFunction{
-			Name:      "exec",
-			Arguments: args(`{"expr": "a < b && c > d"}`),
-		},
-	}
-	if !reflect.DeepEqual(toolCall, expected) {
-		t.Fatalf("expected %#v, got %#v", expected, toolCall)
-	}
-}

model/parsers/lfm2.go

@@ -1,498 +0,0 @@
-package parsers
-
-import (
-	"encoding/json"
-	"errors"
-	"log/slog"
-	"strconv"
-	"strings"
-	"unicode"
-
-	"github.com/ollama/ollama/api"
-)
-
-type LFM2ParserState int
-
-const (
-	LFM2CollectingThinking LFM2ParserState = iota
-	LFM2CollectingContent
-	LFM2CollectingToolCalls
-)
-
-const (
-	lfm2ThinkingOpenTag  = "<think>"
-	lfm2ThinkingCloseTag = "</think>"
-	lfm2ToolCallStartTag = "<|tool_call_start|>"
-	lfm2ToolCallEndTag   = "<|tool_call_end|>"
-)
-
-type LFM2Parser struct {
-	state                    LFM2ParserState
-	buffer                   strings.Builder
-	hasThinkingSupport       bool
-	needsThinkingLeadingTrim bool // trim leading whitespace after <think> tag
-	needsContentLeadingTrim  bool // trim leading whitespace after </think> tag
-}
-
-func (p *LFM2Parser) HasToolSupport() bool {
-	return true
-}
-
-func (p *LFM2Parser) HasThinkingSupport() bool {
-	return p.hasThinkingSupport
-}
-
-func (p *LFM2Parser) setInitialState(lastMessage *api.Message, thinkValue *api.ThinkValue) {
-	prefill := lastMessage != nil && lastMessage.Role == "assistant"
-
-	// Check both model capability AND request preference
-	thinkingEnabled := p.HasThinkingSupport() && (thinkValue != nil && thinkValue.Bool())
-
-	if !thinkingEnabled {
-		p.state = LFM2CollectingContent
-		return
-	}
-
-	if prefill && lastMessage.Content != "" {
-		p.state = LFM2CollectingContent
-		return
-	}
-
-	p.state = LFM2CollectingThinking
-	p.needsThinkingLeadingTrim = true
-}
-
-func (p *LFM2Parser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
-	p.setInitialState(lastMessage, thinkValue)
-	return tools
-}
-
-type lfm2Event interface {
-	isLFM2Event()
-}
-
-type lfm2EventThinkingContent struct {
-	content string
-}
-
-type lfm2EventContent struct {
-	content string
-}
-
-type lfm2EventToolCall struct {
-	toolCall api.ToolCall
-}
-
-func (lfm2EventThinkingContent) isLFM2Event() {}
-func (lfm2EventContent) isLFM2Event()         {}
-func (lfm2EventToolCall) isLFM2Event()        {}
-
-func (p *LFM2Parser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
-	p.buffer.WriteString(s)
-	events := p.parseEvents()
-
-	var toolCalls []api.ToolCall
-	var contentSb strings.Builder
-	var thinkingSb strings.Builder
-	for _, event := range events {
-		switch event := event.(type) {
-		case lfm2EventToolCall:
-			toolCalls = append(toolCalls, event.toolCall)
-		case lfm2EventThinkingContent:
-			thinkingSb.WriteString(event.content)
-		case lfm2EventContent:
-			contentSb.WriteString(event.content)
-		}
-	}
-
-	return contentSb.String(), thinkingSb.String(), toolCalls, nil
-}
-
-func (p *LFM2Parser) parseEvents() []lfm2Event {
-	var all []lfm2Event
-
-	keepLooping := true
-	for keepLooping {
-		var events []lfm2Event
-		events, keepLooping = p.eat()
-		if len(events) > 0 {
-			all = append(all, events...)
-		}
-	}
-
-	return all
-}
-
-func (p *LFM2Parser) eat() ([]lfm2Event, bool) {
-	var events []lfm2Event
-	bufStr := p.buffer.String()
-	if bufStr == "" {
-		return events, false
-	}
-
-	switch p.state {
-	case LFM2CollectingThinking:
-		// Strip opening <think> tag if present
-		if strings.HasPrefix(bufStr, lfm2ThinkingOpenTag) {
-			bufStr = bufStr[len(lfm2ThinkingOpenTag):]
-			p.needsThinkingLeadingTrim = true
-			p.buffer.Reset()
-			p.buffer.WriteString(bufStr)
-		}
-
-		// Trim leading whitespace after <think> tag (may span multiple chunks)
-		if p.needsThinkingLeadingTrim {
-			if trimmed := strings.TrimLeftFunc(bufStr, unicode.IsSpace); trimmed != bufStr {
-				bufStr = trimmed
-				p.buffer.Reset()
-				p.buffer.WriteString(bufStr)
-			}
-			// Clear flag once we have non-whitespace content or buffer is empty
-			if len(bufStr) > 0 {
-				p.needsThinkingLeadingTrim = false
-			}
-		}
-
-		if strings.Contains(bufStr, lfm2ThinkingCloseTag) { // thinking[</think>] -> content
-			split := strings.SplitN(bufStr, lfm2ThinkingCloseTag, 2)
-			thinking := split[0]
-			thinking = strings.TrimRightFunc(thinking, unicode.IsSpace)
-
-			remaining := split[1]
-			remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
-
-			p.buffer.Reset()
-			p.buffer.WriteString(remaining)
-			p.state = LFM2CollectingContent
-			p.needsThinkingLeadingTrim = false
-			// Set flag to trim any additional whitespace that may arrive in later chunks
-			p.needsContentLeadingTrim = len(remaining) == 0
-
-			if len(thinking) > 0 {
-				events = append(events, lfm2EventThinkingContent{content: thinking})
-			}
-			return events, true
-		} else if overlapLen := overlap(bufStr, lfm2ThinkingCloseTag); overlapLen > 0 { // partial </think>
-			beforePartialTag := bufStr[:len(bufStr)-overlapLen]
-			trailingLen := trailingWhitespaceLen(beforePartialTag)
-			ambiguousStart := len(beforePartialTag) - trailingLen
-
-			unambiguous := bufStr[:ambiguousStart]
-			ambiguous := bufStr[ambiguousStart:]
-			p.buffer.Reset()
-			p.buffer.WriteString(ambiguous)
-			if len(unambiguous) > 0 {
-				events = append(events, lfm2EventThinkingContent{content: unambiguous})
-			}
-			return events, false
-		} else { // otherwise its thinking content
-			whitespaceLen := trailingWhitespaceLen(bufStr)
-			ambiguousStart := len(bufStr) - whitespaceLen
-
-			unambiguous := bufStr[:ambiguousStart]
-			ambiguous := bufStr[ambiguousStart:]
-			p.buffer.Reset()
-			p.buffer.WriteString(ambiguous)
-			if len(unambiguous) > 0 {
-				events = append(events, lfm2EventThinkingContent{content: unambiguous})
-			}
-			return events, false
-		}
-
-	case LFM2CollectingContent:
-		// Trim leading whitespace after </think> tag (may span multiple chunks)
-		if p.needsContentLeadingTrim {
-			if trimmed := strings.TrimLeftFunc(bufStr, unicode.IsSpace); trimmed != bufStr {
-				bufStr = trimmed
-				p.buffer.Reset()
-				p.buffer.WriteString(bufStr)
-			}
-			// Clear flag once we have non-whitespace content
-			if len(bufStr) > 0 {
-				p.needsContentLeadingTrim = false
-			}
-		}
-
-		if strings.Contains(bufStr, lfm2ToolCallStartTag) { // content[<|tool_call_start|>] -> tool calls
-			split := strings.SplitN(bufStr, lfm2ToolCallStartTag, 2)
-			contentBefore := strings.TrimRightFunc(split[0], unicode.IsSpace)
-			remaining := split[1]
-
-			p.buffer.Reset()
-			p.buffer.WriteString(remaining)
-			p.state = LFM2CollectingToolCalls
-
-			if len(contentBefore) > 0 {
-				events = append(events, lfm2EventContent{content: contentBefore})
-			}
-			return events, true
-		} else { // otherwise its content
-			p.buffer.Reset()
-			if len(bufStr) > 0 {
-				events = append(events, lfm2EventContent{content: bufStr})
-			}
-			return events, false
-		}
-
-	case LFM2CollectingToolCalls:
-		// Look for complete tool call JSON between tags
-		if idx := strings.Index(bufStr, lfm2ToolCallEndTag); idx != -1 {
-			toolCallContent := bufStr[:idx]
-
-			if toolCalls, err := p.parseToolCallsContent(toolCallContent); err == nil && len(toolCalls) > 0 {
-				remaining := bufStr[idx+len(lfm2ToolCallEndTag):]
-
-				// Check if there's another tool call
-				if strings.HasPrefix(remaining, lfm2ToolCallStartTag) {
-					remaining = remaining[len(lfm2ToolCallStartTag):]
-				} else {
-					// No more tool calls, go back to content
-					remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
-					p.state = LFM2CollectingContent
-				}
-
-				p.buffer.Reset()
-				p.buffer.WriteString(remaining)
-
-				for _, tc := range toolCalls {
-					events = append(events, lfm2EventToolCall{toolCall: tc})
-				}
-				return events, true
-			} else if err != nil {
-				slog.Warn("lfm2 tool call parsing failed", "error", err, "content", toolCallContent)
-			}
-		}
-
-		return events, false
-	}
-
-	return events, false
-}
-
-// parseToolCallsContent parses one or more tool calls from content
-// Supports JSON format and Python-style format including multiple calls: [func1(...),func2(...)]
-func (p *LFM2Parser) parseToolCallsContent(content string) ([]api.ToolCall, error) {
-	content = strings.TrimSpace(content)
-
-	// Try JSON format first: {"name": "func", "arguments": {...}}
-	var parsed struct {
-		Name      string          `json:"name"`
-		Arguments json.RawMessage `json:"arguments"`
-	}
-
-	if err := json.Unmarshal([]byte(content), &parsed); err == nil && parsed.Name != "" {
-		var args api.ToolCallFunctionArguments
-		if len(parsed.Arguments) > 0 {
-			if err := json.Unmarshal(parsed.Arguments, &args); err != nil {
-				return nil, err
-			}
-		} else {
-			args = api.NewToolCallFunctionArguments()
-		}
-
-		return []api.ToolCall{{
-			Function: api.ToolCallFunction{
-				Name:      parsed.Name,
-				Arguments: args,
-			},
-		}}, nil
-	}
-
-	// Try Python-style format: [func(arg1='val1'),func2(arg2='val2')] or func(arg1='val1')
-	return p.parsePythonStyleToolCalls(content)
-}
-
-// parsePythonStyleToolCalls parses one or more Python-style tool calls
-// Examples: [bash(command='ls'),bash(command='pwd')] or bash(command='ls')
-func (p *LFM2Parser) parsePythonStyleToolCalls(content string) ([]api.ToolCall, error) {
-	content = strings.TrimSpace(content)
-
-	// Strip outer brackets if present: [func(...)] -> func(...)
-	if strings.HasPrefix(content, "[") && strings.HasSuffix(content, "]") {
-		content = content[1 : len(content)-1]
-	}
-
-	var toolCalls []api.ToolCall
-
-	// Parse multiple function calls separated by commas at the top level
-	for len(content) > 0 {
-		content = strings.TrimSpace(content)
-		if content == "" {
-			break
-		}
-
-		// Skip leading comma from previous iteration
-		if strings.HasPrefix(content, ",") {
-			content = strings.TrimSpace(content[1:])
-			if content == "" {
-				break
-			}
-		}
-
-		// Find function name
-		parenIdx := strings.Index(content, "(")
-		if parenIdx == -1 {
-			return nil, errors.New("invalid tool call: no opening parenthesis")
-		}
-
-		funcName := strings.TrimSpace(content[:parenIdx])
-		if funcName == "" {
-			return nil, errors.New("invalid tool call: empty function name")
-		}
-
-		// Find matching closing parenthesis
-		closeIdx := findMatchingParen(content, parenIdx)
-		if closeIdx == -1 {
-			return nil, errors.New("invalid tool call: no matching closing parenthesis")
-		}
-
-		argsStr := content[parenIdx+1 : closeIdx]
-		args := api.NewToolCallFunctionArguments()
-
-		if argsStr != "" {
-			if err := parsePythonArgs(argsStr, &args); err != nil {
-				return nil, err
-			}
-		}
-
-		toolCalls = append(toolCalls, api.ToolCall{
-			Function: api.ToolCallFunction{
-				Name:      funcName,
-				Arguments: args,
-			},
-		})
-
-		// Move past this function call
-		content = content[closeIdx+1:]
-	}
-
-	if len(toolCalls) == 0 {
-		return nil, errors.New("no tool calls found")
-	}
-
-	return toolCalls, nil
-}
-
-// findMatchingParen finds the index of the closing parenthesis matching the one at openIdx
-// Returns -1 if not found. Handles nested parentheses and quoted strings.
-func findMatchingParen(s string, openIdx int) int {
-	depth := 1
-	i := openIdx + 1
-	for i < len(s) && depth > 0 {
-		switch s[i] {
-		case '(':
-			depth++
-		case ')':
-			depth--
-			if depth == 0 {
-				return i
-			}
-		case '\'', '"':
-			// Skip quoted string
-			quote := s[i]
-			i++
-			for i < len(s) && s[i] != quote {
-				if s[i] == '\\' && i+1 < len(s) {
-					i++ // skip escaped char
-				}
-				i++
-			}
-		}
-		i++
-	}
-	return -1
-}
-
-// parseToolCallContent parses a single tool call (for backward compatibility with tests)
-func (p *LFM2Parser) parseToolCallContent(content string) (api.ToolCall, error) {
-	calls, err := p.parseToolCallsContent(content)
-	if err != nil {
-		return api.ToolCall{}, err
-	}
-	if len(calls) == 0 {
-		return api.ToolCall{}, errors.New("no tool call found")
-	}
-	return calls[0], nil
-}
-
-// parsePythonArgs parses Python-style keyword arguments: key='value', key2="value2"
-func parsePythonArgs(argsStr string, args *api.ToolCallFunctionArguments) error {
-	// Simple state machine to parse key='value' pairs
-	// Handles: command='ls', flag="-la", count=42, enabled=true
-	var key string
-	i := 0
-
-	for i < len(argsStr) {
-		// Skip whitespace
-		for i < len(argsStr) && (argsStr[i] == ' ' || argsStr[i] == '\t' || argsStr[i] == '\n') {
-			i++
-		}
-		if i >= len(argsStr) {
-			break
-		}
-
-		// Parse key
-		keyStart := i
-		for i < len(argsStr) && argsStr[i] != '=' && argsStr[i] != ',' {
-			i++
-		}
-		if i >= len(argsStr) || argsStr[i] != '=' {
-			return errors.New("invalid argument: expected '='")
-		}
-		key = strings.TrimSpace(argsStr[keyStart:i])
-		i++ // skip '='
-
-		// Skip whitespace after =
-		for i < len(argsStr) && (argsStr[i] == ' ' || argsStr[i] == '\t') {
-			i++
-		}
-
-		// Parse value
-		var value string
-		if i < len(argsStr) && (argsStr[i] == '\'' || argsStr[i] == '"') {
-			// Quoted string
-			quote := argsStr[i]
-			i++
-			valueStart := i
-			for i < len(argsStr) && argsStr[i] != quote {
-				if argsStr[i] == '\\' && i+1 < len(argsStr) {
-					i += 2 // skip escaped char
-				} else {
-					i++
-				}
-			}
-			value = argsStr[valueStart:i]
-			if i < len(argsStr) {
-				i++ // skip closing quote
-			}
-			args.Set(key, value)
-		} else {
-			// Unquoted value (number, bool, etc)
-			valueStart := i
-			for i < len(argsStr) && argsStr[i] != ',' {
-				i++
-			}
-			value = strings.TrimSpace(argsStr[valueStart:i])
-
-			// Try to parse as number or bool
-			if v, err := strconv.ParseInt(value, 10, 64); err == nil {
-				args.Set(key, v)
-			} else if v, err := strconv.ParseFloat(value, 64); err == nil {
-				args.Set(key, v)
-			} else if value == "true" {
-				args.Set(key, true)
-			} else if value == "false" {
-				args.Set(key, false)
-			} else {
-				args.Set(key, value)
-			}
-		}
-
-		// Skip comma and whitespace
-		for i < len(argsStr) && (argsStr[i] == ',' || argsStr[i] == ' ' || argsStr[i] == '\t' || argsStr[i] == '\n') {
-			i++
-		}
-	}
-
-	return nil
-}

model/parsers/nemotron3nano.go

@@ -1,6 +1,7 @@
 package parsers
 
 import (
+	"regexp"
 	"strings"
 	"unicode"
 
@@ -13,114 +14,243 @@ const (
 	Nemotron3NanoCollectingThinking Nemotron3NanoParserState = iota
 	Nemotron3NanoSkipWhitespaceAfterThinking
 	Nemotron3NanoCollectingContent
+	Nemotron3NanoCollectingToolCalls
 )
 
 const (
-	nemotronThinkClose   = "</think>"
-	nemotronToolCallOpen = "<tool_call>"
+	nemotronThinkClose    = "</think>"
+	nemotronToolCallOpen  = "<tool_call>"
+	nemotronToolCallClose = "</tool_call>"
 )
 
 type Nemotron3NanoParser struct {
-	state      Nemotron3NanoParserState
-	buffer     strings.Builder
-	toolParser *Qwen3CoderParser
+	state  Nemotron3NanoParserState
+	buffer strings.Builder
+	tools  []api.Tool
 }
 
 func (p *Nemotron3NanoParser) HasToolSupport() bool     { return true }
 func (p *Nemotron3NanoParser) HasThinkingSupport() bool { return true }
 
 func (p *Nemotron3NanoParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
-	p.toolParser = &Qwen3CoderParser{}
-	p.toolParser.Init(tools, nil, nil)
+	p.tools = tools
 
+	// thinking is enabled if user requests it
 	thinkingEnabled := thinkValue != nil && thinkValue.Bool()
+
 	prefill := lastMessage != nil && lastMessage.Role == "assistant"
 
-	if !thinkingEnabled || (prefill && lastMessage.Content != "") {
+	if !thinkingEnabled {
 		p.state = Nemotron3NanoCollectingContent
-	} else {
-		p.state = Nemotron3NanoCollectingThinking
+		return tools
 	}
 
+	if prefill && lastMessage.Content != "" {
+		p.state = Nemotron3NanoCollectingContent
+		return tools
+	}
+
+	p.state = Nemotron3NanoCollectingThinking
 	return tools
 }
 
+type nemotronEvent interface {
+	isNemotronEvent()
+}
+
+type nemotronEventThinkingContent struct {
+	content string
+}
+
+type nemotronEventContent struct {
+	content string
+}
+
+type nemotronEventToolCall struct {
+	toolCall api.ToolCall
+}
+
+func (nemotronEventThinkingContent) isNemotronEvent() {}
+func (nemotronEventContent) isNemotronEvent()         {}
+func (nemotronEventToolCall) isNemotronEvent()        {}
+
 func (p *Nemotron3NanoParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
-	if p.state == Nemotron3NanoCollectingContent {
-		return p.toolParser.Add(s, done)
+	p.buffer.WriteString(s)
+	events := p.parseEvents()
+
+	var toolCalls []api.ToolCall
+	var contentSb strings.Builder
+	var thinkingSb strings.Builder
+	for _, event := range events {
+		switch event := event.(type) {
+		case nemotronEventToolCall:
+			toolCalls = append(toolCalls, event.toolCall)
+		case nemotronEventThinkingContent:
+			thinkingSb.WriteString(event.content)
+		case nemotronEventContent:
+			contentSb.WriteString(event.content)
+		}
 	}
 
-	if p.state == Nemotron3NanoSkipWhitespaceAfterThinking {
-		s = strings.TrimLeftFunc(s, unicode.IsSpace)
-		if s == "" {
-			return "", "", nil, nil
+	return contentSb.String(), thinkingSb.String(), toolCalls, nil
+}
+
+func (p *Nemotron3NanoParser) parseEvents() []nemotronEvent {
+	var all []nemotronEvent
+
+	keepLooping := true
+	for keepLooping {
+		var events []nemotronEvent
+		events, keepLooping = p.eat()
+		if len(events) > 0 {
+			all = append(all, events...)
+		}
+	}
+
+	return all
+}
+
+// emitWithPartialCheck extracts unambiguous content before a potential partial tag
+func (p *Nemotron3NanoParser) emitWithPartialCheck(bufStr, tag string) (unambiguous, ambiguous string) {
+	if overlapLen := overlap(bufStr, tag); overlapLen > 0 {
+		beforePartialTag := bufStr[:len(bufStr)-overlapLen]
+		trailingLen := trailingWhitespaceLen(beforePartialTag)
+		return bufStr[:len(beforePartialTag)-trailingLen], bufStr[len(beforePartialTag)-trailingLen:]
+	}
+	wsLen := trailingWhitespaceLen(bufStr)
+	return bufStr[:len(bufStr)-wsLen], bufStr[len(bufStr)-wsLen:]
+}
+
+func (p *Nemotron3NanoParser) eat() ([]nemotronEvent, bool) {
+	bufStr := p.buffer.String()
+	if bufStr == "" {
+		return nil, false
+	}
+
+	switch p.state {
+	case Nemotron3NanoCollectingThinking:
+		if strings.Contains(bufStr, nemotronThinkClose) {
+			split := strings.SplitN(bufStr, nemotronThinkClose, 2)
+			thinking := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			p.buffer.Reset()
+			remainder := strings.TrimLeftFunc(split[1], unicode.IsSpace)
+			p.buffer.WriteString(remainder)
+			// Transition to whitespace-skipping state if buffer is empty,
+			// otherwise go directly to content collection
+			if remainder == "" {
+				p.state = Nemotron3NanoSkipWhitespaceAfterThinking
+			} else {
+				p.state = Nemotron3NanoCollectingContent
+			}
+			if thinking != "" {
+				return []nemotronEvent{nemotronEventThinkingContent{content: thinking}}, true
+			}
+			return nil, true
+		}
+		unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronThinkClose)
+		p.buffer.Reset()
+		p.buffer.WriteString(ambig)
+		if unambig != "" {
+			return []nemotronEvent{nemotronEventThinkingContent{content: unambig}}, false
+		}
+		return nil, false
+
+	// We only want to skip whitespace between thinking and content
+	case Nemotron3NanoSkipWhitespaceAfterThinking:
+		bufStr = strings.TrimLeftFunc(bufStr, unicode.IsSpace)
+		p.buffer.Reset()
+		p.buffer.WriteString(bufStr)
+		if bufStr == "" {
+			return nil, false
 		}
 		p.state = Nemotron3NanoCollectingContent
-		return p.toolParser.Add(s, done)
-	}
+		return nil, true
 
-	// Nemotron3NanoCollectingThinking - buffer and look for end markers
-	p.buffer.WriteString(s)
-	bufStr := p.buffer.String()
-
-	// Look for end of thinking: </think> or <tool_call> (model may skip </think>)
-	thinkIdx := strings.Index(bufStr, nemotronThinkClose)
-	toolIdx := strings.Index(bufStr, nemotronToolCallOpen)
-
-	var endIdx int = -1
-	var remainder string
-
-	if thinkIdx != -1 && (toolIdx == -1 || thinkIdx < toolIdx) {
-		endIdx = thinkIdx
-		remainder = strings.TrimLeftFunc(bufStr[thinkIdx+len(nemotronThinkClose):], unicode.IsSpace)
-	} else if toolIdx != -1 {
-		endIdx = toolIdx
-		remainder = bufStr[toolIdx:] // Include <tool_call> tag
-	}
-
-	if endIdx != -1 {
-		thinking = strings.TrimRightFunc(bufStr[:endIdx], unicode.IsSpace)
-		p.buffer.Reset()
-
-		if remainder == "" {
-			p.state = Nemotron3NanoSkipWhitespaceAfterThinking
-		} else {
-			p.state = Nemotron3NanoCollectingContent
-			content, _, calls, err = p.toolParser.Add(remainder, done)
+	case Nemotron3NanoCollectingContent:
+		if strings.Contains(bufStr, nemotronToolCallOpen) {
+			split := strings.SplitN(bufStr, nemotronToolCallOpen, 2)
+			content := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			p.buffer.Reset()
+			p.buffer.WriteString(split[1])
+			p.state = Nemotron3NanoCollectingToolCalls
+			if content != "" {
+				return []nemotronEvent{nemotronEventContent{content: content}}, true
+			}
+			return nil, true
 		}
-		return content, thinking, calls, err
+		unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronToolCallOpen)
+		p.buffer.Reset()
+		p.buffer.WriteString(ambig)
+		if unambig != "" {
+			return []nemotronEvent{nemotronEventContent{content: unambig}}, false
+		}
+		return nil, false
+
+	case Nemotron3NanoCollectingToolCalls:
+		if strings.Contains(bufStr, nemotronToolCallClose) {
+			split := strings.SplitN(bufStr, nemotronToolCallClose, 2)
+			remaining := strings.TrimLeftFunc(split[1], unicode.IsSpace)
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+
+			var events []nemotronEvent
+			if tc, err := p.parseToolCall(split[0]); err == nil {
+				events = append(events, nemotronEventToolCall{toolCall: tc})
+			}
+
+			if !strings.Contains(remaining, nemotronToolCallOpen) {
+				p.state = Nemotron3NanoCollectingContent
+			}
+			return events, true
+		}
+		return nil, false
 	}
 
-	// No end marker - emit unambiguous thinking
-	thinking = p.emitThinking(bufStr)
-	return "", thinking, nil, nil
+	return nil, false
 }
 
-// emitThinking returns unambiguous thinking content, keeping potential partial tags in buffer
-func (p *Nemotron3NanoParser) emitThinking(bufStr string) string {
-	// Check for partial </think> or <tool_call> at end
-	thinkOverlap := overlap(bufStr, nemotronThinkClose)
-	toolOverlap := overlap(bufStr, nemotronToolCallOpen)
-	maxOverlap := max(thinkOverlap, toolOverlap)
+var (
+	nemotronFunctionRegex  = regexp.MustCompile(`<function=([^>]+)>`)
+	nemotronParameterRegex = regexp.MustCompile(`<parameter=([^>]+)>\n?([\s\S]*?)\n?</parameter>`)
+)
 
-	if maxOverlap > 0 {
-		unambiguous := bufStr[:len(bufStr)-maxOverlap]
-		unambiguous = strings.TrimRightFunc(unambiguous, unicode.IsSpace)
-		p.buffer.Reset()
-		p.buffer.WriteString(bufStr[len(bufStr)-maxOverlap:])
-		return unambiguous
+func (p *Nemotron3NanoParser) parseToolCall(content string) (api.ToolCall, error) {
+	toolCall := api.ToolCall{}
+
+	// Extract function name
+	fnMatch := nemotronFunctionRegex.FindStringSubmatch(content)
+	if len(fnMatch) < 2 {
+		return toolCall, nil
+	}
+	toolCall.Function.Name = fnMatch[1]
+
+	// Extract parameters
+	toolCall.Function.Arguments = api.NewToolCallFunctionArguments()
+	paramMatches := nemotronParameterRegex.FindAllStringSubmatch(content, -1)
+	for _, match := range paramMatches {
+		if len(match) >= 3 {
+			paramName := match[1]
+			paramValue := strings.TrimSpace(match[2])
+
+			// Try to parse as typed value based on tool definition
+			toolCall.Function.Arguments.Set(paramName, p.parseParamValue(paramName, paramValue))
+		}
 	}
 
-	// No partial tags - emit all but trailing whitespace
-	wsLen := trailingWhitespaceLen(bufStr)
-	if wsLen > 0 {
-		unambiguous := bufStr[:len(bufStr)-wsLen]
-		p.buffer.Reset()
-		p.buffer.WriteString(bufStr[len(bufStr)-wsLen:])
-		return unambiguous
-	}
-
-	// Nothing to hold back
-	p.buffer.Reset()
-	return bufStr
+	return toolCall, nil
+}
+
+func (p *Nemotron3NanoParser) parseParamValue(paramName string, raw string) any {
+	// Find the matching tool to get parameter type
+	var paramType api.PropertyType
+	for _, tool := range p.tools {
+		if tool.Function.Parameters.Properties != nil {
+			if prop, ok := tool.Function.Parameters.Properties.Get(paramName); ok {
+				paramType = prop.Type
+				break
+			}
+		}
+	}
+
+	return parseValue(raw, paramType)
 }

model/parsers/nemotron3nano_test.go

@@ -8,8 +8,6 @@ import (
 	"github.com/ollama/ollama/api"
 )
 
-// TestNemotron3NanoParser tests Nemotron-specific behavior (thinking support).
-// Tool call parsing is tested in qwen3coder_test.go since Nemotron delegates to Qwen3CoderParser.
 func TestNemotron3NanoParser(t *testing.T) {
 	tests := []struct {
 		name             string
@@ -19,6 +17,18 @@ func TestNemotron3NanoParser(t *testing.T) {
 		expectedThinking string
 		expectedCalls    []api.ToolCall
 	}{
+		{
+			name:            "simple content - no thinking",
+			input:           "Hello, how can I help you?",
+			thinkValue:      nil,
+			expectedContent: "Hello, how can I help you?",
+		},
+		{
+			name:            "simple content - thinking disabled",
+			input:           "Hello, how can I help you?",
+			thinkValue:      &api.ThinkValue{Value: false},
+			expectedContent: "Hello, how can I help you?",
+		},
 		{
 			name:             "thinking then content",
 			input:            "Let me think about this...</think>\nHere is my answer.",
@@ -33,6 +43,69 @@ func TestNemotron3NanoParser(t *testing.T) {
 			expectedThinking: "Step 1: Analyze\nStep 2: Process\nStep 3: Conclude",
 			expectedContent:  "The answer is 42.",
 		},
+		{
+			name:       "simple tool call",
+			input:      "<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"city": "Paris"}),
+					},
+				},
+			},
+		},
+		{
+			name:            "content then tool call",
+			input:           "Let me check the weather.\n<tool_call>\n<function=get_weather>\n<parameter=city>\nNYC\n</parameter>\n</function>\n</tool_call>",
+			thinkValue:      nil,
+			expectedContent: "Let me check the weather.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"city": "NYC"}),
+					},
+				},
+			},
+		},
+		{
+			name:       "tool call with multiple parameters",
+			input:      "<tool_call>\n<function=book_flight>\n<parameter=from>\nSFO\n</parameter>\n<parameter=to>\nNYC\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "book_flight",
+						Arguments: testArgs(map[string]any{
+							"from": "SFO",
+							"to":   "NYC",
+						}),
+					},
+				},
+			},
+		},
+		{
+			name: "multiple tool calls",
+			input: "<tool_call>\n<function=get_weather>\n<parameter=city>\nSan Francisco\n</parameter>\n</function>\n</tool_call>\n" +
+				"<tool_call>\n<function=get_weather>\n<parameter=city>\nNew York\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"city": "San Francisco"}),
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"city": "New York"}),
+					},
+				},
+			},
+		},
 		{
 			name:             "thinking then tool call",
 			input:            "I should check the weather...</think>\n<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>",
@@ -62,6 +135,19 @@ func TestNemotron3NanoParser(t *testing.T) {
 				},
 			},
 		},
+		{
+			name:       "tool call with multiline parameter value",
+			input:      "<tool_call>\n<function=create_note>\n<parameter=content>\nLine 1\nLine 2\nLine 3\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "create_note",
+						Arguments: testArgs(map[string]any{"content": "Line 1\nLine 2\nLine 3"}),
+					},
+				},
+			},
+		},
 		{
 			name:             "empty thinking block - immediate close",
 			input:            "</think>\nHere is my answer.",
@@ -75,6 +161,18 @@ func TestNemotron3NanoParser(t *testing.T) {
 			thinkValue:      &api.ThinkValue{Value: false},
 			expectedContent: "</think>\nSome content after spurious tag.",
 		},
+		{
+			name:          "tool call with no function name - returns empty tool call",
+			input:         "<tool_call>\n<function=>\n</function>\n</tool_call>",
+			thinkValue:    nil,
+			expectedCalls: []api.ToolCall{{Function: api.ToolCallFunction{Name: "", Arguments: api.NewToolCallFunctionArguments()}}},
+		},
+		{
+			name:            "content with newlines preserved",
+			input:           "Line 1\n\nLine 2\n\n\nLine 3",
+			thinkValue:      nil,
+			expectedContent: "Line 1\n\nLine 2\n\n\nLine 3",
+		},
 		{
 			name:             "thinking with only whitespace after close tag",
 			input:            "My thoughts...</think>   \n\t\n   Content here.",
@@ -82,6 +180,25 @@ func TestNemotron3NanoParser(t *testing.T) {
 			expectedThinking: "My thoughts...",
 			expectedContent:  "Content here.",
 		},
+		{
+			name:            "unicode content",
+			input:           "Hello 世界! 🌍 Ñoño",
+			thinkValue:      nil,
+			expectedContent: "Hello 世界! 🌍 Ñoño",
+		},
+		{
+			name:       "tool call with numeric parameter",
+			input:      "<tool_call>\n<function=set_temp>\n<parameter=value>\n42\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "set_temp",
+						Arguments: testArgs(map[string]any{"value": "42"}),
+					},
+				},
+			},
+		},
 	}
 
 	for _, tt := range tests {
@@ -116,8 +233,6 @@ func TestNemotron3NanoParser(t *testing.T) {
 	}
 }
 
-// TestNemotron3NanoParser_Streaming tests streaming behavior for thinking support.
-// Tool call streaming is tested in qwen3coder_test.go.
 func TestNemotron3NanoParser_Streaming(t *testing.T) {
 	tests := []struct {
 		name             string
@@ -127,6 +242,18 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 		expectedThinking string
 		expectedCalls    []api.ToolCall
 	}{
+		{
+			name:            "streaming content character by character",
+			chunks:          []string{"H", "e", "l", "l", "o", ",", " ", "w", "o", "r", "l", "d", "!"},
+			thinkValue:      nil,
+			expectedContent: "Hello, world!",
+		},
+		{
+			name:            "streaming content small tokens",
+			chunks:          []string{"Hel", "lo", ", ", "how ", "can", " I", " help", " you", " today", "?"},
+			thinkValue:      nil,
+			expectedContent: "Hello, how can I help you today?",
+		},
 		{
 			name:             "streaming thinking then content - granular",
 			chunks:           []string{"Let", " me", " th", "ink", " about", " this", "...", "<", "/", "think", ">", "\n", "Here", " is", " my", " answer", "."},
@@ -141,6 +268,45 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 			expectedThinking: "Step 1: Analyze\nStep 2: Process",
 			expectedContent:  "The answer.",
 		},
+		{
+			name:       "streaming tool call - highly granular",
+			chunks:     []string{"<", "tool", "_", "call", ">", "\n", "<", "func", "tion", "=", "get", "_", "weather", ">", "\n", "<", "param", "eter", "=", "city", ">", "\n", "Par", "is", "\n", "</", "param", "eter", ">", "\n", "</", "func", "tion", ">", "\n", "</", "tool", "_", "call", ">"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"city": "Paris"}),
+					},
+				},
+			},
+		},
+		{
+			name:            "streaming content then tool call - granular",
+			chunks:          []string{"Let", " me", " check", " the", " weather", ".", "\n<", "tool_call", ">", "\n", "<function=", "get_weather", ">", "\n", "<parameter=", "city", ">", "\n", "NYC", "\n", "</parameter>", "\n", "</function>", "\n", "</tool_call>"},
+			thinkValue:      nil,
+			expectedContent: "Let me check the weather.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"city": "NYC"}),
+					},
+				},
+			},
+		},
+		{
+			name:   "tool call tag split character by character",
+			chunks: []string{"<", "t", "o", "o", "l", "_", "c", "a", "l", "l", ">", "\n", "<", "f", "u", "n", "c", "t", "i", "o", "n", "=", "t", "e", "s", "t", ">", "\n", "<", "/", "f", "u", "n", "c", "t", "i", "o", "n", ">", "\n", "<", "/", "t", "o", "o", "l", "_", "c", "a", "l", "l", ">"},
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "test",
+						Arguments: api.NewToolCallFunctionArguments(),
+					},
+				},
+			},
+		},
 		{
 			name:             "thinking close tag split character by character",
 			chunks:           []string{"I", "'", "m", " ", "t", "h", "i", "n", "k", "i", "n", "g", ".", ".", ".", "<", "/", "t", "h", "i", "n", "k", ">", "\n", "D", "o", "n", "e", "!"},
@@ -155,6 +321,22 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 			expectedThinking: "Thinking...",
 			expectedContent:  "Content here.",
 		},
+		{
+			name:       "tool call with multiple parameters - streaming",
+			chunks:     []string{"<tool_", "call>\n", "<function", "=book_", "flight>", "\n<para", "meter=", "from>\n", "SFO\n", "</param", "eter>", "\n<param", "eter=to", ">\nNYC", "\n</para", "meter>", "\n</func", "tion>\n", "</tool_", "call>"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "book_flight",
+						Arguments: testArgs(map[string]any{
+							"from": "SFO",
+							"to":   "NYC",
+						}),
+					},
+				},
+			},
+		},
 		{
 			name:             "thinking then content then tool call - streaming",
 			chunks:           []string{"Ana", "lyzing", " your", " request", "...", "</", "think", ">\n", "I'll", " check", " that", " for", " you", ".", "\n", "<tool", "_call", ">\n", "<function", "=search", ">\n", "<parameter", "=query", ">\n", "test", " query", "\n</", "parameter", ">\n", "</function", ">\n", "</tool", "_call", ">"},
@@ -170,6 +352,45 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 				},
 			},
 		},
+		{
+			name: "multiple tool calls - streaming",
+			chunks: []string{
+				"<tool_call>", "\n", "<function=", "get_weather>", "\n",
+				"<parameter=", "city>\n", "San Fran", "cisco\n", "</parameter>", "\n",
+				"</function>", "\n", "</tool_call>", "\n",
+				"<tool_", "call>\n", "<function", "=get_weather", ">\n",
+				"<param", "eter=city", ">\nNew", " York\n", "</parameter>\n",
+				"</function>\n", "</tool_call>",
+			},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"city": "San Francisco"}),
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"city": "New York"}),
+					},
+				},
+			},
+		},
+		{
+			name:       "tool call with multiline parameter - streaming",
+			chunks:     []string{"<tool_call>\n", "<function=", "create_note>\n", "<parameter=", "content>\n", "Line 1", "\nLine", " 2\n", "Line 3", "\n</parameter>\n", "</function>\n", "</tool_call>"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "create_note",
+						Arguments: testArgs(map[string]any{"content": "Line 1\nLine 2\nLine 3"}),
+					},
+				},
+			},
+		},
 		{
 			name:             "empty thinking block",
 			chunks:           []string{"</think>", "\n", "Just content."},
@@ -177,6 +398,12 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 			expectedThinking: "",
 			expectedContent:  "Just content.",
 		},
+		{
+			name:            "empty input chunks interspersed",
+			chunks:          []string{"Hello", "", " ", "", "world", "", "!"},
+			thinkValue:      nil,
+			expectedContent: "Hello world!",
+		},
 		{
 			name:             "tool call immediately after think close - no content",
 			chunks:           []string{"Analyzing...", "</think>", "\n", "<tool_call>", "\n<function=test>\n</function>\n", "</tool_call>"},
@@ -191,6 +418,25 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
 				},
 			},
 		},
+		{
+			name:       "tool call with empty parameter value",
+			chunks:     []string{"<tool_call>\n<function=test>\n<parameter=name>\n", "\n</parameter>\n</function>\n</tool_call>"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "test",
+						Arguments: testArgs(map[string]any{"name": ""}),
+					},
+				},
+			},
+		},
+		{
+			name:            "partial tool call tag at end - buffered",
+			chunks:          []string{"Here's some content", "<tool"},
+			thinkValue:      nil,
+			expectedContent: "Here's some content",
+		},
 	}
 
 	for _, tt := range tests {
@@ -326,65 +572,3 @@ func TestNemotron3NanoParser_WithTools(t *testing.T) {
 		t.Errorf("calls mismatch (-got +want):\n%s", diff)
 	}
 }
-
-// TestNemotron3NanoParser_ToolCallWithoutThinkClose tests the case where thinking is enabled
-// but the model outputs content + tool call WITHOUT the </think> tag.
-// The parser should still parse the tool call (content before is treated as thinking).
-func TestNemotron3NanoParser_ToolCallWithoutThinkClose(t *testing.T) {
-	chunks := []string{
-		"Let", " me", " analyze", " this", ".", "\n",
-		"<tool_call>", "\n",
-		"<function=get_weather>", "\n",
-		"<parameter=city>", "Paris", "</parameter>", "\n",
-		"</function>", "\n",
-		"</tool_call>",
-	}
-
-	p := &Nemotron3NanoParser{}
-	p.Init(nil, nil, &api.ThinkValue{Value: true}) // thinking ENABLED but model doesn't output </think>
-
-	var allContent string
-	var allThinking string
-	var allCalls []api.ToolCall
-
-	for _, chunk := range chunks {
-		content, thinking, calls, err := p.Add(chunk, false)
-		if err != nil {
-			t.Fatalf("unexpected error: %v", err)
-		}
-		allContent += content
-		allThinking += thinking
-		allCalls = append(allCalls, calls...)
-	}
-
-	// Drain
-	content, thinking, calls, err := p.Add("", true)
-	if err != nil {
-		t.Fatalf("unexpected error on done: %v", err)
-	}
-	allContent += content
-	allThinking += thinking
-	allCalls = append(allCalls, calls...)
-
-	// The parser was in thinking mode, so text before <tool_call> is emitted as thinking.
-	expectedThinking := "Let me analyze this."
-
-	expectedCalls := []api.ToolCall{
-		{
-			Function: api.ToolCallFunction{
-				Name:      "get_weather",
-				Arguments: testArgs(map[string]any{"city": "Paris"}),
-			},
-		},
-	}
-
-	if allContent != "" {
-		t.Errorf("expected no content (text was streamed as thinking), got: %q", allContent)
-	}
-	if diff := cmp.Diff(allThinking, expectedThinking); diff != "" {
-		t.Errorf("thinking mismatch (-got +want):\n%s", diff)
-	}
-	if diff := cmp.Diff(allCalls, expectedCalls, argsComparer); diff != "" {
-		t.Errorf("calls mismatch (-got +want):\n%s", diff)
-	}
-}

model/parsers/parsers.go

@@ -68,12 +68,6 @@ func ParserForName(name string) Parser {
 		return &Nemotron3NanoParser{}
 	case "functiongemma":
 		return &FunctionGemmaParser{}
-	case "glm-4.7":
-		return &GLM47Parser{}
-	case "lfm2":
-		return &LFM2Parser{hasThinkingSupport: false}
-	case "lfm2-thinking":
-		return &LFM2Parser{hasThinkingSupport: true}
 	default:
 		return nil
 	}

model/parsers/qwen3coder_test.go

@@ -91,37 +91,6 @@ func TestQwenParserStreaming(t *testing.T) {
 				},
 			},
 		},
-		{
-			desc: "tool call tags split character by character",
-			steps: []step{
-				{input: "<", wantEvents: []qwenEvent{}},
-				{input: "t", wantEvents: []qwenEvent{}},
-				{input: "o", wantEvents: []qwenEvent{}},
-				{input: "o", wantEvents: []qwenEvent{}},
-				{input: "l", wantEvents: []qwenEvent{}},
-				{input: "_", wantEvents: []qwenEvent{}},
-				{input: "c", wantEvents: []qwenEvent{}},
-				{input: "a", wantEvents: []qwenEvent{}},
-				{input: "l", wantEvents: []qwenEvent{}},
-				{input: "l", wantEvents: []qwenEvent{}},
-				{input: ">", wantEvents: []qwenEvent{}},
-				{input: "a", wantEvents: []qwenEvent{}},
-				{input: "b", wantEvents: []qwenEvent{}},
-				{input: "c", wantEvents: []qwenEvent{}},
-				{input: "<", wantEvents: []qwenEvent{}},
-				{input: "/", wantEvents: []qwenEvent{}},
-				{input: "t", wantEvents: []qwenEvent{}},
-				{input: "o", wantEvents: []qwenEvent{}},
-				{input: "o", wantEvents: []qwenEvent{}},
-				{input: "l", wantEvents: []qwenEvent{}},
-				{input: "_", wantEvents: []qwenEvent{}},
-				{input: "c", wantEvents: []qwenEvent{}},
-				{input: "a", wantEvents: []qwenEvent{}},
-				{input: "l", wantEvents: []qwenEvent{}},
-				{input: "l", wantEvents: []qwenEvent{}},
-				{input: ">", wantEvents: []qwenEvent{qwenEventRawToolCall{raw: "abc"}}},
-			},
-		},
 		{
 			desc: "trailing whitespace between content and tool call",
 			steps: []step{

model/parsers/testhelpers_test.go

@@ -96,11 +96,3 @@ func testArgs(m map[string]any) api.ToolCallFunctionArguments {
 	}
 	return args
 }
-
-func args(s string) api.ToolCallFunctionArguments {
-	var result api.ToolCallFunctionArguments
-	if err := json.Unmarshal([]byte(s), &result); err != nil {
-		panic("invalid JSON in args(): " + err.Error())
-	}
-	return result
-}

model/renderers/glm46.go

@@ -1,110 +0,0 @@
-package renderers
-
-import (
-	"encoding/json"
-	"fmt"
-	"strings"
-
-	"github.com/ollama/ollama/api"
-)
-
-type GLM46Renderer struct{}
-
-func (r *GLM46Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
-	var sb strings.Builder
-
-	sb.WriteString("[gMASK]<sop>")
-
-	var lastUserIndex int
-	for i, message := range messages {
-		if message.Role == "user" {
-			lastUserIndex = i
-		}
-	}
-
-	if len(tools) > 0 {
-		sb.WriteString("<|system|>\n")
-		sb.WriteString("# Tools\n\n")
-		sb.WriteString("You may call one or more functions to assist with the user query.\n\n")
-		sb.WriteString("You are provided with function signatures within <tools></tools> XML tags:\n")
-		sb.WriteString("<tools>\n")
-		for _, tool := range tools {
-			d, _ := json.Marshal(tool)
-			sb.WriteString(string(d) + "\n")
-		}
-		sb.WriteString("</tools>\n\n")
-		sb.WriteString("For each function call, output the function name and arguments within the following XML format:\n")
-		sb.WriteString("<tool_call>{function-name}\n")
-		sb.WriteString("<arg_key>{arg-key-1}</arg_key>\n")
-		sb.WriteString("<arg_value>{arg-value-1}</arg_value>\n")
-		sb.WriteString("<arg_key>{arg-key-2}</arg_key>\n")
-		sb.WriteString("<arg_value>{arg-value-2}</arg_value>\n")
-		sb.WriteString("...\n")
-		sb.WriteString("</tool_call>")
-	}
-
-	for i, message := range messages {
-		switch message.Role {
-		case "user":
-			sb.WriteString("<|user|>\n")
-			sb.WriteString(message.Content)
-			if thinkValue != nil && !thinkValue.Bool() && !strings.HasSuffix(message.Content, "/nothink") {
-				sb.WriteString("/nothink")
-			}
-		case "assistant":
-			sb.WriteString("<|assistant|>")
-			if i > lastUserIndex {
-				if message.Thinking != "" {
-					sb.WriteString("\n<think>" + message.Thinking + "</think>")
-				} else {
-					sb.WriteString("\n<think></think>")
-				}
-			}
-			if message.Content != "" {
-				sb.WriteString("\n" + message.Content)
-			}
-			if len(message.ToolCalls) > 0 {
-				for _, toolCall := range message.ToolCalls {
-					sb.WriteString("\n<tool_call>" + toolCall.Function.Name + "\n")
-					for key, value := range toolCall.Function.Arguments.All() {
-						sb.WriteString("<arg_key>" + key + "</arg_key>\n")
-
-						var valueStr string
-						if str, ok := value.(string); ok {
-							valueStr = str
-						} else {
-							jsonBytes, err := json.Marshal(value)
-							if err != nil {
-								valueStr = fmt.Sprintf("%v", value)
-							} else {
-								valueStr = string(jsonBytes)
-							}
-						}
-
-						sb.WriteString("<arg_value>" + valueStr + "</arg_value>\n")
-					}
-
-					sb.WriteString("</tool_call>")
-				}
-			}
-		case "tool":
-			if i == 0 || messages[i-1].Role != "tool" {
-				sb.WriteString("<|observation|>")
-			}
-			sb.WriteString("\n<tool_response>\n")
-			sb.WriteString(message.Content)
-			sb.WriteString("\n</tool_response>")
-		case "system":
-			sb.WriteString("<|system|>\n")
-			sb.WriteString(message.Content)
-		}
-	}
-
-	// Add generation prompt
-	sb.WriteString("<|assistant|>")
-	if thinkValue != nil && !thinkValue.Bool() {
-		sb.WriteString("\n<think></think>\n")
-	}
-
-	return sb.String(), nil
-}

model/renderers/glm46_test.go

@@ -1,223 +0,0 @@
-package renderers
-
-import (
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-	"github.com/ollama/ollama/api"
-)
-
-func TestGLM46Renderer(t *testing.T) {
-	tests := []struct {
-		name       string
-		messages   []api.Message
-		tools      []api.Tool
-		thinkValue *api.ThinkValue
-		expected   string
-		skip       string
-	}{
-		{
-			name: "basic",
-			messages: []api.Message{
-				{Role: "user", Content: "Hello, how are you?"},
-			},
-			expected: `[gMASK]<sop><|user|>
-Hello, how are you?<|assistant|>`,
-		},
-		{
-			name: "basic with system message",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant."},
-				{Role: "user", Content: "Hello, how are you?"},
-			},
-			expected: `[gMASK]<sop><|system|>
-You are a helpful assistant.<|user|>
-Hello, how are you?<|assistant|>`,
-		},
-		{
-			name: "basic with user assistant user",
-			messages: []api.Message{
-				{Role: "user", Content: "What is the capital of France?"},
-				{Role: "assistant", Thinking: "Let me analyze the request...", Content: "The capital of France is Paris."},
-				{Role: "user", Content: "Fantastic!"},
-			},
-			expected: `[gMASK]<sop><|user|>
-What is the capital of France?<|assistant|>
-The capital of France is Paris.<|user|>
-Fantastic!<|assistant|>`,
-		},
-		{
-			skip: "tool call ordering not guaranteed yet",
-			name: "tools",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant with access to tools."},
-				{Role: "user", Content: "What is the weather like in Tokyo?"},
-			},
-			tools: []api.Tool{
-				{
-					Type: "function",
-					Function: api.ToolFunction{
-						Name:        "get_weather",
-						Description: "Get the current weather in a given location",
-						Parameters: api.ToolFunctionParameters{
-							Type:       "object",
-							Required:   []string{"location"},
-							Properties: propsMap(`{"location": {"type": "string", "description": "The city and state, e.g. San Francisco, CA"}, "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}}`),
-						},
-					},
-				},
-			},
-			expected: `[gMASK]<sop><|system|>
-# Tools
-
-You may call one or more functions to assist with the user query.
-
-You are provided with function signatures within <tools></tools> XML tags:
-<tools>
-{"type":"function","function":{"name":"get_weather","description":"Get the current weather in a given location","parameters":{"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"},"unit":{"type":"string","description":"","enum":["celsius","fahrenheit"]}}}}}
-</tools>
-
-For each function call, output the function name and arguments within the following XML format:
-<tool_call>{function-name}
-<arg_key>{arg-key-1}</arg_key>
-<arg_value>{arg-value-1}</arg_value>
-<arg_key>{arg-key-2}</arg_key>
-<arg_value>{arg-value-2}</arg_value>
-...
-</tool_call><|system|>
-You are a helpful assistant with access to tools.<|user|>
-What is the weather like in Tokyo?<|assistant|>`,
-		},
-		{
-			skip: "tool call ordering not guaranteed yet",
-			name: "tool calls",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant with access to tools."},
-				{Role: "user", Content: "What is the weather like in Tokyo?"},
-				{
-					Role: "assistant",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name:      "get_weather",
-								Arguments: args(`{"location": "Tokyo, Japan", "unit": "celsius"}`),
-							},
-						},
-						{
-							Function: api.ToolCallFunction{
-								Name:      "get_weather",
-								Arguments: args(`{"location": "Japan", "unit": "fahrenheit"}`),
-							},
-						},
-					},
-				},
-				{
-					Role:     "tool",
-					Content:  "{\"temperature\": 22, \"weather\": \"partly cloudy\", \"humidity\": 65}",
-					ToolName: "get_weather",
-				},
-				{
-					Role:     "tool",
-					Content:  "{\"temperature\": 68, \"weather\": \"sunny\", \"humidity\": 75}",
-					ToolName: "get_weather",
-				},
-				{
-					Role:    "assistant",
-					Content: "The weather in Tokyo is currently partly cloudy with a temperature of 22°C and 65% humidity. It's a pleasant day with moderate temperatures.",
-				},
-			},
-			tools: []api.Tool{
-				{
-					Type: "function",
-					Function: api.ToolFunction{
-						Name:        "get_weather",
-						Description: "Get the current weather in a given location",
-						Parameters: api.ToolFunctionParameters{
-							Type:       "object",
-							Required:   []string{"location"},
-							Properties: propsMap(`{"location": {"type": "string", "description": "The city and state, e.g. San Francisco, CA"}, "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}}`),
-						},
-					},
-				},
-			},
-			expected: `[gMASK]<sop><|system|>
-# Tools
-
-You may call one or more functions to assist with the user query.
-
-You are provided with function signatures within <tools></tools> XML tags:
-<tools>
-{"type":"function","function":{"name":"get_weather","description":"Get the current weather in a given location","parameters":{"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"},"unit":{"type":"string","description":"","enum":["celsius","fahrenheit"]}}}}}
-</tools>
-
-For each function call, output the function name and arguments within the following XML format:
-<tool_call>{function-name}
-<arg_key>{arg-key-1}</arg_key>
-<arg_value>{arg-value-1}</arg_value>
-<arg_key>{arg-key-2}</arg_key>
-<arg_value>{arg-value-2}</arg_value>
-...
-</tool_call><|system|>
-You are a helpful assistant with access to tools.<|user|>
-What is the weather like in Tokyo?<|assistant|>
-<think></think>
-<tool_call>get_weather
-<arg_key>location</arg_key>
-<arg_value>Tokyo, Japan</arg_value>
-<arg_key>unit</arg_key>
-<arg_value>celsius</arg_value>
-</tool_call>
-<tool_call>get_weather
-<arg_key>location</arg_key>
-<arg_value>Japan</arg_value>
-<arg_key>unit</arg_key>
-<arg_value>fahrenheit</arg_value>
-</tool_call><|observation|>
-<tool_response>
-{"temperature": 22, "weather": "partly cloudy", "humidity": 65}
-</tool_response>
-<tool_response>
-{"temperature": 68, "weather": "sunny", "humidity": 75}
-</tool_response><|assistant|>
-<think></think>
-The weather in Tokyo is currently partly cloudy with a temperature of 22°C and 65% humidity. It's a pleasant day with moderate temperatures.<|assistant|>`,
-		},
-		{
-			name: "think true",
-			messages: []api.Message{
-				{Role: "user", Content: "Hello, how are you?"},
-			},
-			thinkValue: &api.ThinkValue{Value: true},
-			expected: `[gMASK]<sop><|user|>
-Hello, how are you?<|assistant|>`,
-		},
-		{
-			name: "think false",
-			messages: []api.Message{
-				{Role: "user", Content: "Hello, how are you?"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected: `[gMASK]<sop><|user|>
-Hello, how are you?/nothink<|assistant|>
-<think></think>
-`,
-		},
-	}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			if tt.skip != "" {
-				t.Skip(tt.skip)
-			}
-			renderer := &GLM46Renderer{}
-			rendered, err := renderer.Render(tt.messages, tt.tools, tt.thinkValue)
-			if err != nil {
-				t.Fatal(err)
-			}
-			if diff := cmp.Diff(rendered, tt.expected); diff != "" {
-				t.Errorf("mismatch (-got +want):\n%s", diff)
-				t.Logf("Got:\n%s", rendered)
-				t.Logf("Expected:\n%s", tt.expected)
-			}
-		})
-	}
-}

model/renderers/glm47.go

@@ -1,170 +0,0 @@
-package renderers
-
-import (
-	"encoding/json"
-	"fmt"
-	"strings"
-
-	"github.com/ollama/ollama/api"
-)
-
-// GLM47Renderer renders messages for GLM-4.7 models.
-//
-// GLM-4.7 Thinking Modes (ref: https://docs.z.ai/guides/capabilities/thinking-mode):
-//
-//  1. INTERLEAVED THINKING
-//     The model thinks between tool calls and after receiving tool results.
-//     This enables complex step-by-step reasoning: interpreting each tool output
-//     before deciding what to do next. Thinking blocks are preserved and returned
-//     with tool results to maintain reasoning continuity.
-//
-//  2. PRESERVED THINKING
-//     The model retains reasoning content from previous assistant turns in context.
-//     This preserves reasoning continuity across multi-turn conversations. The
-//     upstream API has a "clear_thinking" parameter to control this:
-//     - clear_thinking=true:  clears reasoning from previous turns (outputs </think>)
-//     - clear_thinking=false: preserves <think>...</think> blocks from previous turns
-//
-//  3. TURN-LEVEL THINKING
-//     Controls whether the model should reason on each turn. The upstream API
-//     uses "enable_thinking" parameter:
-//     - enable_thinking=true:  outputs <think> to start reasoning
-//     - enable_thinking=false: outputs </think> to skip reasoning
-//
-// OLLAMA DEFAULTS:
-//   - Thinking is ENABLED by default (thinkValue=nil or true outputs <think>)
-//   - Thinking is PRESERVED by default (reasoning content from previous turns is always
-//     included in <think>...</think> blocks, equivalent to clear_thinking=false)
-//   - Users can disable thinking per-turn via thinkValue=false
-type GLM47Renderer struct{}
-
-func (r *GLM47Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
-	var sb strings.Builder
-
-	sb.WriteString("[gMASK]<sop>")
-
-	if len(tools) > 0 {
-		sb.WriteString("<|system|>\n")
-		sb.WriteString("# Tools\n\n")
-		sb.WriteString("You may call one or more functions to assist with the user query.\n\n")
-		sb.WriteString("You are provided with function signatures within <tools></tools> XML tags:\n")
-		sb.WriteString("<tools>\n")
-		for _, tool := range tools {
-			d, _ := json.Marshal(tool)
-			sb.WriteString(formatGLM47ToolJSON(d))
-			sb.WriteString("\n")
-		}
-		sb.WriteString("</tools>\n\n")
-		sb.WriteString("For each function call, output the function name and arguments within the following XML format:\n")
-		sb.WriteString("<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call>")
-	}
-
-	think := true
-	if thinkValue != nil && !thinkValue.Bool() {
-		think = false
-	}
-
-	for i, message := range messages {
-		switch message.Role {
-		case "user":
-			sb.WriteString("<|user|>")
-			sb.WriteString(message.Content)
-		case "assistant":
-			sb.WriteString("<|assistant|>")
-			if message.Thinking != "" {
-				sb.WriteString("<think>" + message.Thinking + "</think>")
-			} else {
-				sb.WriteString("</think>")
-			}
-			if message.Content != "" {
-				sb.WriteString(message.Content)
-			}
-			if len(message.ToolCalls) > 0 {
-				for _, toolCall := range message.ToolCalls {
-					sb.WriteString("<tool_call>" + toolCall.Function.Name)
-					sb.WriteString(renderGLM47ToolArguments(toolCall.Function.Arguments))
-					sb.WriteString("</tool_call>")
-				}
-			}
-		case "tool":
-			if i == 0 || messages[i-1].Role != "tool" {
-				sb.WriteString("<|observation|>")
-			}
-			sb.WriteString("<tool_response>")
-			sb.WriteString(message.Content)
-			sb.WriteString("</tool_response>")
-		case "system":
-			sb.WriteString("<|system|>")
-			sb.WriteString(message.Content)
-		}
-	}
-
-	sb.WriteString("<|assistant|>")
-	if think {
-		sb.WriteString("<think>")
-	} else {
-		sb.WriteString("</think>")
-	}
-
-	return sb.String(), nil
-}
-
-func renderGLM47ToolArguments(args api.ToolCallFunctionArguments) string {
-	var sb strings.Builder
-	for key, value := range args.All() {
-		sb.WriteString("<arg_key>" + key + "</arg_key>")
-		var valueStr string
-		if str, ok := value.(string); ok {
-			valueStr = str
-		} else {
-			jsonBytes, err := json.Marshal(value)
-			if err != nil {
-				valueStr = fmt.Sprintf("%v", value)
-			} else {
-				valueStr = string(jsonBytes)
-			}
-		}
-
-		sb.WriteString("<arg_value>" + valueStr + "</arg_value>")
-	}
-
-	return sb.String()
-}
-
-func formatGLM47ToolJSON(raw []byte) string {
-	var sb strings.Builder
-	sb.Grow(len(raw) + len(raw)/10)
-
-	inString := false
-	escaped := false
-	for i := range raw {
-		ch := raw[i]
-		sb.WriteByte(ch)
-
-		if inString {
-			if escaped {
-				escaped = false
-				continue
-			}
-			if ch == '\\' {
-				escaped = true
-				continue
-			}
-			if ch == '"' {
-				inString = false
-			}
-			continue
-		}
-
-		if ch == '"' {
-			inString = true
-			continue
-		}
-
-		if ch == ':' || ch == ',' {
-			sb.WriteByte(' ')
-		}
-	}
-
-	return sb.String()
-}

model/renderers/glm47_test.go

@@ -1,191 +0,0 @@
-package renderers
-
-import (
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-	"github.com/ollama/ollama/api"
-)
-
-func TestGLM47Renderer(t *testing.T) {
-	tests := []struct {
-		name       string
-		messages   []api.Message
-		tools      []api.Tool
-		thinkValue *api.ThinkValue
-		expected   string
-	}{
-		{
-			name: "basic user message",
-			messages: []api.Message{
-				{Role: "user", Content: "Hello"},
-			},
-			expected: "[gMASK]<sop><|user|>Hello<|assistant|><think>",
-		},
-		{
-			name: "thinking disabled",
-			messages: []api.Message{
-				{Role: "user", Content: "Hello"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "[gMASK]<sop><|user|>Hello<|assistant|></think>",
-		},
-		{
-			name: "system and user",
-			messages: []api.Message{
-				{Role: "system", Content: "You are helpful."},
-				{Role: "user", Content: "Hello"},
-			},
-			expected: "[gMASK]<sop><|system|>You are helpful.<|user|>Hello<|assistant|><think>",
-		},
-		{
-			name: "multi-turn conversation",
-			messages: []api.Message{
-				{Role: "user", Content: "Hi"},
-				{Role: "assistant", Content: "Hello there"},
-				{Role: "user", Content: "How are you?"},
-			},
-			expected: "[gMASK]<sop><|user|>Hi<|assistant|></think>Hello there<|user|>How are you?<|assistant|><think>",
-		},
-		{
-			name: "assistant with reasoning_content",
-			messages: []api.Message{
-				{Role: "user", Content: "Answer with reasoning."},
-				{Role: "assistant", Thinking: "Plan.", Content: "Done."},
-			},
-			expected: "[gMASK]<sop><|user|>Answer with reasoning.<|assistant|><think>Plan.</think>Done.<|assistant|><think>",
-		},
-		{
-			name: "tool call with empty content",
-			messages: []api.Message{
-				{Role: "user", Content: "Weather?"},
-				{
-					Role: "assistant",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name:      "get_weather",
-								Arguments: args(`{"location": "Tokyo", "unit": "celsius"}`),
-							},
-						},
-					},
-				},
-				{Role: "tool", Content: `{"temperature":22}`},
-			},
-			tools: []api.Tool{
-				{
-					Type: "function",
-					Function: api.ToolFunction{
-						Name:        "get_weather",
-						Description: "Get weather",
-						Parameters: api.ToolFunctionParameters{
-							Type:       "object",
-							Required:   []string{"location"},
-							Properties: propsMap(`{"location": {"type": "string"}}`),
-						},
-					},
-				},
-			},
-			expected: "[gMASK]<sop><|system|>\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"description\": \"Get weather\", \"parameters\": {\"type\": \"object\", \"required\": [\"location\"], \"properties\": {\"location\": {\"type\": \"string\"}}}}}\n</tools>\n\nFor each function call, output the function name and arguments within the following XML format:\n<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call><|user|>Weather?<|assistant|></think><tool_call>get_weather<arg_key>location</arg_key><arg_value>Tokyo</arg_value><arg_key>unit</arg_key><arg_value>celsius</arg_value></tool_call><|observation|><tool_response>{\"temperature\":22}</tool_response><|assistant|><think>",
-		},
-		{
-			name: "tool call with content",
-			messages: []api.Message{
-				{Role: "user", Content: "Weather?"},
-				{
-					Role:    "assistant",
-					Content: "Let me check",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name:      "get_weather",
-								Arguments: args(`{"location": "Tokyo"}`),
-							},
-						},
-					},
-				},
-				{Role: "tool", Content: `{"temperature":22}`},
-				{Role: "assistant", Content: "It is 22C."},
-			},
-			tools: []api.Tool{
-				{
-					Type: "function",
-					Function: api.ToolFunction{
-						Name:        "get_weather",
-						Description: "Get weather",
-						Parameters: api.ToolFunctionParameters{
-							Type:       "object",
-							Required:   []string{"location"},
-							Properties: propsMap(`{"location": {"type": "string"}}`),
-						},
-					},
-				},
-			},
-			expected: "[gMASK]<sop><|system|>\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"description\": \"Get weather\", \"parameters\": {\"type\": \"object\", \"required\": [\"location\"], \"properties\": {\"location\": {\"type\": \"string\"}}}}}\n</tools>\n\nFor each function call, output the function name and arguments within the following XML format:\n<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call><|user|>Weather?<|assistant|></think>Let me check<tool_call>get_weather<arg_key>location</arg_key><arg_value>Tokyo</arg_value></tool_call><|observation|><tool_response>{\"temperature\":22}</tool_response><|assistant|></think>It is 22C.<|assistant|><think>",
-		},
-		{
-			name: "multiple tool calls and responses",
-			messages: []api.Message{
-				{Role: "user", Content: "Compare weather"},
-				{
-					Role: "assistant",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name:      "get_weather",
-								Arguments: args(`{"location": "Tokyo"}`),
-							},
-						},
-						{
-							Function: api.ToolCallFunction{
-								Name:      "get_weather",
-								Arguments: args(`{"location": "Paris"}`),
-							},
-						},
-					},
-				},
-				{Role: "tool", Content: `{"temperature":22}`},
-				{Role: "tool", Content: `{"temperature":18}`},
-			},
-			tools: []api.Tool{
-				{
-					Type: "function",
-					Function: api.ToolFunction{
-						Name:        "get_weather",
-						Description: "Get weather",
-						Parameters: api.ToolFunctionParameters{
-							Type:       "object",
-							Required:   []string{"location"},
-							Properties: propsMap(`{"location": {"type": "string"}}`),
-						},
-					},
-				},
-			},
-			expected: "[gMASK]<sop><|system|>\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"description\": \"Get weather\", \"parameters\": {\"type\": \"object\", \"required\": [\"location\"], \"properties\": {\"location\": {\"type\": \"string\"}}}}}\n</tools>\n\nFor each function call, output the function name and arguments within the following XML format:\n<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call><|user|>Compare weather<|assistant|></think><tool_call>get_weather<arg_key>location</arg_key><arg_value>Tokyo</arg_value></tool_call><tool_call>get_weather<arg_key>location</arg_key><arg_value>Paris</arg_value></tool_call><|observation|><tool_response>{\"temperature\":22}</tool_response><tool_response>{\"temperature\":18}</tool_response><|assistant|><think>",
-		},
-		{
-			name: "preserved thinking in multi-turn",
-			messages: []api.Message{
-				{Role: "user", Content: "Think step by step"},
-				{Role: "assistant", Thinking: "Let me think...", Content: "Here's my answer."},
-				{Role: "user", Content: "Continue"},
-			},
-			expected: "[gMASK]<sop><|user|>Think step by step<|assistant|><think>Let me think...</think>Here's my answer.<|user|>Continue<|assistant|><think>",
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			renderer := &GLM47Renderer{}
-			rendered, err := renderer.Render(tt.messages, tt.tools, tt.thinkValue)
-			if err != nil {
-				t.Fatal(err)
-			}
-			if diff := cmp.Diff(rendered, tt.expected); diff != "" {
-				t.Errorf("mismatch (-got +want):\n%s", diff)
-				t.Logf("Got:\n%s", rendered)
-				t.Logf("Expected:\n%s", tt.expected)
-			}
-		})
-	}
-}

model/renderers/lfm2.go

@@ -1,144 +0,0 @@
-package renderers
-
-import (
-	"encoding/json"
-	"strings"
-
-	"github.com/ollama/ollama/api"
-)
-
-type LFM2Renderer struct {
-	IsThinking bool
-}
-
-func (r *LFM2Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
-	var sb strings.Builder
-
-	// Note: BOS token is added by the tokenizer (add_bos_token: true), not the renderer
-
-	// Extract first system message if present (to combine with tools)
-	var firstSystemContent string
-	startIdx := 0
-	if len(messages) > 0 && messages[0].Role == "system" {
-		firstSystemContent = messages[0].Content
-		startIdx = 1
-	}
-
-	// Append tools to first system content
-	if len(tools) > 0 {
-		if firstSystemContent != "" {
-			firstSystemContent += "\n"
-		}
-		firstSystemContent += "List of tools: ["
-		for i, tool := range tools {
-			toolJSON, err := json.Marshal(tool)
-			if err != nil {
-				return "", err
-			}
-			firstSystemContent += string(toolJSON)
-			if i < len(tools)-1 {
-				firstSystemContent += ", "
-			}
-		}
-		firstSystemContent += "]"
-	}
-
-	// Output first system block if it has content
-	if firstSystemContent != "" {
-		sb.WriteString("<|im_start|>system\n")
-		sb.WriteString(firstSystemContent)
-		sb.WriteString("<|im_end|>\n")
-	}
-
-	// Find the index of the last assistant message for thinking stripping
-	lastAssistantIndex := -1
-	for i := len(messages) - 1; i >= startIdx; i-- {
-		if messages[i].Role == "assistant" {
-			lastAssistantIndex = i
-			break
-		}
-	}
-
-	// Track whether we need to add generation prompt
-	needsGenerationPrompt := len(messages) > 0
-
-	for i := startIdx; i < len(messages); i++ {
-		message := messages[i]
-		switch message.Role {
-		case "system":
-			// Additional system messages (after the first) are rendered normally
-			sb.WriteString("<|im_start|>system\n")
-			sb.WriteString(message.Content)
-			sb.WriteString("<|im_end|>\n")
-
-		case "user":
-			sb.WriteString("<|im_start|>user\n")
-			sb.WriteString(message.Content)
-			sb.WriteString("<|im_end|>\n")
-			needsGenerationPrompt = true
-
-		case "assistant":
-			sb.WriteString("<|im_start|>assistant\n")
-
-			// Check if this is the last assistant message
-			isLastAssistant := i == lastAssistantIndex
-
-			// Process content (may need thinking stripped)
-			content := message.Content
-
-			// Handle thinking tags in assistant content
-			keepPastThinking := r.IsThinking && (thinkValue != nil && thinkValue.Bool())
-			if strings.Contains(content, "</think>") {
-				parts := strings.SplitN(content, "</think>", 2)
-				if len(parts) > 1 {
-					if !isLastAssistant && !keepPastThinking {
-						// Strip thinking entirely for past assistant messages
-						content = strings.TrimSpace(parts[1])
-					} else {
-						// Preserve thinking but trim whitespace after </think>
-						content = parts[0] + "</think>" + strings.TrimLeft(parts[1], " \t\n\r")
-					}
-				}
-			}
-
-			if len(message.ToolCalls) > 0 {
-				// Assistant with tool calls - write content first (if any after stripping)
-				if content != "" {
-					sb.WriteString(content)
-				}
-
-				for _, toolCall := range message.ToolCalls {
-					sb.WriteString("<|tool_call_start|>")
-					toolCallJSON := map[string]any{
-						"name":      toolCall.Function.Name,
-						"arguments": toolCall.Function.Arguments,
-					}
-					callJSON, _ := json.Marshal(toolCallJSON)
-					sb.WriteString(string(callJSON))
-					sb.WriteString("<|tool_call_end|>")
-				}
-			} else {
-				sb.WriteString(content)
-			}
-
-			sb.WriteString("<|im_end|>\n")
-			needsGenerationPrompt = true // Always add gen prompt after assistant when add_generation_prompt=true
-
-		case "tool":
-			// Tool responses are rendered as plain messages per the chat template
-			sb.WriteString("<|im_start|>tool\n")
-			sb.WriteString(message.Content)
-			sb.WriteString("<|im_end|>\n")
-			needsGenerationPrompt = true
-		}
-	}
-
-	// Add generation prompt
-	if needsGenerationPrompt {
-		sb.WriteString("<|im_start|>assistant\n")
-		// Note: Model is a "thinking-only" model - it will output <think> itself
-		// We don't add <think> tag to the prompt
-	}
-
-	return sb.String(), nil
-}

model/renderers/lfm2_test.go

@@ -1,427 +0,0 @@
-package renderers
-
-import (
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-
-	"github.com/ollama/ollama/api"
-)
-
-func TestLFM2Renderer(t *testing.T) {
-	tests := []struct {
-		name       string
-		messages   []api.Message
-		tools      []api.Tool
-		thinkValue *api.ThinkValue
-		expected   string
-	}{
-		{
-			name: "basic user message",
-			messages: []api.Message{
-				{Role: "user", Content: "Hello!"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "basic with system message",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant."},
-				{Role: "user", Content: "Hello!"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "multiple system messages rendered separately",
-			messages: []api.Message{
-				{Role: "system", Content: "First instruction."},
-				{Role: "system", Content: "Second instruction."},
-				{Role: "user", Content: "Hello!"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>system\nFirst instruction.<|im_end|>\n<|im_start|>system\nSecond instruction.<|im_end|>\n<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "multi-turn conversation",
-			messages: []api.Message{
-				{Role: "user", Content: "What is 2+2?"},
-				{Role: "assistant", Content: "The answer is 4."},
-				{Role: "user", Content: "Thanks!"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nWhat is 2+2?<|im_end|>\n<|im_start|>assistant\nThe answer is 4.<|im_end|>\n<|im_start|>user\nThanks!<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "only system message",
-			messages: []api.Message{
-				{Role: "system", Content: "You are helpful."},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>system\nYou are helpful.<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			// When assistant is the LAST assistant, thinking is preserved (even with keep_past_thinking=false)
-			name: "user-assistant-user: last assistant preserves thinking",
-			messages: []api.Message{
-				{Role: "user", Content: "Q1"},
-				{Role: "assistant", Content: "<think>reasoning</think>A1"},
-				{Role: "user", Content: "Q2"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nQ1<|im_end|>\n<|im_start|>assistant\n<think>reasoning</think>A1<|im_end|>\n<|im_start|>user\nQ2<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			// With two assistants, first is stripped (not last), second preserved (is last)
-			name: "multi-turn thinking: first stripped, second preserved",
-			messages: []api.Message{
-				{Role: "user", Content: "Q1"},
-				{Role: "assistant", Content: "<think>reason1</think>A1"},
-				{Role: "user", Content: "Q2"},
-				{Role: "assistant", Content: "<think>reason2</think>A2"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nQ1<|im_end|>\n<|im_start|>assistant\nA1<|im_end|>\n<|im_start|>user\nQ2<|im_end|>\n<|im_start|>assistant\n<think>reason2</think>A2<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			// With thinking enabled (keep_past_thinking=true), both preserved
-			name: "multi-turn thinking: both preserved when thinking enabled",
-			messages: []api.Message{
-				{Role: "user", Content: "Q1"},
-				{Role: "assistant", Content: "<think>reason1</think>A1"},
-				{Role: "user", Content: "Q2"},
-				{Role: "assistant", Content: "<think>reason2</think>A2"},
-			},
-			thinkValue: &api.ThinkValue{Value: true},
-			expected:   "<|im_start|>user\nQ1<|im_end|>\n<|im_start|>assistant\n<think>reason1</think>A1<|im_end|>\n<|im_start|>user\nQ2<|im_end|>\n<|im_start|>assistant\n<think>reason2</think>A2<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "assistant with tool calls",
-			messages: []api.Message{
-				{Role: "user", Content: "What's the weather?"},
-				{
-					Role: "assistant",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name: "get_weather",
-								Arguments: testArgs(map[string]any{
-									"location": "Paris",
-								}),
-							},
-						},
-					},
-				},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   `<|im_start|>user` + "\n" + `What's the weather?<|im_end|>` + "\n" + `<|im_start|>assistant` + "\n" + `<|tool_call_start|>{"arguments":{"location":"Paris"},"name":"get_weather"}<|tool_call_end|><|im_end|>` + "\n" + `<|im_start|>assistant` + "\n",
-		},
-		{
-			name: "assistant with content and tool calls",
-			messages: []api.Message{
-				{Role: "user", Content: "What's the weather in Paris?"},
-				{
-					Role:    "assistant",
-					Content: "Let me check.",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name: "get_weather",
-								Arguments: testArgs(map[string]any{
-									"location": "Paris",
-								}),
-							},
-						},
-					},
-				},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   `<|im_start|>user` + "\n" + `What's the weather in Paris?<|im_end|>` + "\n" + `<|im_start|>assistant` + "\n" + `Let me check.<|tool_call_start|>{"arguments":{"location":"Paris"},"name":"get_weather"}<|tool_call_end|><|im_end|>` + "\n" + `<|im_start|>assistant` + "\n",
-		},
-		{
-			name: "tool response",
-			messages: []api.Message{
-				{Role: "user", Content: "What's the weather?"},
-				{Role: "assistant", Content: "Let me check."},
-				{Role: "tool", Content: "22C, Sunny"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nWhat's the weather?<|im_end|>\n<|im_start|>assistant\nLet me check.<|im_end|>\n<|im_start|>tool\n22C, Sunny<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "multiple tool calls",
-			messages: []api.Message{
-				{Role: "user", Content: "Get weather for Paris and London"},
-				{
-					Role: "assistant",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name: "get_weather",
-								Arguments: testArgs(map[string]any{
-									"location": "Paris",
-								}),
-							},
-						},
-						{
-							Function: api.ToolCallFunction{
-								Name: "get_weather",
-								Arguments: testArgs(map[string]any{
-									"location": "London",
-								}),
-							},
-						},
-					},
-				},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   `<|im_start|>user` + "\n" + `Get weather for Paris and London<|im_end|>` + "\n" + `<|im_start|>assistant` + "\n" + `<|tool_call_start|>{"arguments":{"location":"Paris"},"name":"get_weather"}<|tool_call_end|><|tool_call_start|>{"arguments":{"location":"London"},"name":"get_weather"}<|tool_call_end|><|im_end|>` + "\n" + `<|im_start|>assistant` + "\n",
-		},
-		{
-			name: "tools definitions with system message",
-			messages: []api.Message{
-				{Role: "system", Content: "You are helpful."},
-				{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",
-							Properties: testPropsMap(map[string]api.ToolProperty{
-								"location": {
-									Type:        api.PropertyType{"string"},
-									Description: "City name",
-								},
-							}),
-							Required: []string{"location"},
-						},
-					},
-				},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   `<|im_start|>system` + "\n" + `You are helpful.` + "\n" + `List of tools: [{"type":"function","function":{"name":"get_weather","description":"Get current weather","parameters":{"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}}}]<|im_end|>` + "\n" + `<|im_start|>user` + "\n" + `What's the weather?<|im_end|>` + "\n" + `<|im_start|>assistant` + "\n",
-		},
-		{
-			name: "tools definitions without system message",
-			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",
-							Properties: testPropsMap(map[string]api.ToolProperty{
-								"location": {
-									Type:        api.PropertyType{"string"},
-									Description: "City name",
-								},
-							}),
-							Required: []string{"location"},
-						},
-					},
-				},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   `<|im_start|>system` + "\n" + `List of tools: [{"type":"function","function":{"name":"get_weather","description":"Get current weather","parameters":{"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}}}]<|im_end|>` + "\n" + `<|im_start|>user` + "\n" + `What's the weather?<|im_end|>` + "\n" + `<|im_start|>assistant` + "\n",
-		},
-		{
-			name: "multiple tools without system message",
-			messages: []api.Message{
-				{Role: "user", Content: "Hello"},
-			},
-			tools: []api.Tool{
-				{
-					Type: "function",
-					Function: api.ToolFunction{
-						Name:        "get_weather",
-						Description: "Get weather",
-					},
-				},
-				{
-					Type: "function",
-					Function: api.ToolFunction{
-						Name:        "get_time",
-						Description: "Get time",
-					},
-				},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>system\nList of tools: [{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get weather\",\"parameters\":{\"type\":\"\",\"properties\":null}}}, {\"type\":\"function\",\"function\":{\"name\":\"get_time\",\"description\":\"Get time\",\"parameters\":{\"type\":\"\",\"properties\":null}}}]<|im_end|>\n<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "user-tool sequence",
-			messages: []api.Message{
-				{Role: "user", Content: "Check weather"},
-				{Role: "tool", Content: "22C"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nCheck weather<|im_end|>\n<|im_start|>tool\n22C<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "full tool call cycle",
-			messages: []api.Message{
-				{Role: "user", Content: "Check weather"},
-				{Role: "assistant", Content: "Let me check"},
-				{Role: "tool", Content: "22C"},
-				{Role: "assistant", Content: "It's 22C"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nCheck weather<|im_end|>\n<|im_start|>assistant\nLet me check<|im_end|>\n<|im_start|>tool\n22C<|im_end|>\n<|im_start|>assistant\nIt's 22C<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "unicode content",
-			messages: []api.Message{
-				{Role: "user", Content: "你好世界! مرحبا 🌍"},
-				{Role: "assistant", Content: "Hello! 👋"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\n你好世界! مرحبا 🌍<|im_end|>\n<|im_start|>assistant\nHello! 👋<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "newlines in content",
-			messages: []api.Message{
-				{Role: "user", Content: "Line 1\nLine 2\n\nLine 4"},
-				{Role: "assistant", Content: "Response with\nmultiple\nlines"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nLine 1\nLine 2\n\nLine 4<|im_end|>\n<|im_start|>assistant\nResponse with\nmultiple\nlines<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			name: "empty assistant content",
-			messages: []api.Message{
-				{Role: "user", Content: "Hello"},
-				{Role: "assistant", Content: ""},
-				{Role: "user", Content: "OK"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\n<|im_end|>\n<|im_start|>user\nOK<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			// Generation prompt does NOT include <think> - model outputs it
-			name: "generation prompt has no think tag",
-			messages: []api.Message{
-				{Role: "user", Content: "Think hard"},
-			},
-			thinkValue: &api.ThinkValue{Value: true},
-			expected:   "<|im_start|>user\nThink hard<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			// Interleaved: thinking before tool call - last assistant preserves thinking
-			name: "thinking before tool call (last assistant)",
-			messages: []api.Message{
-				{Role: "user", Content: "What's the weather?"},
-				{
-					Role:    "assistant",
-					Content: "<think>I need to check the weather</think>",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name: "get_weather",
-								Arguments: testArgs(map[string]any{
-									"location": "Paris",
-								}),
-							},
-						},
-					},
-				},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nWhat's the weather?<|im_end|>\n<|im_start|>assistant\n<think>I need to check the weather</think><|tool_call_start|>{\"arguments\":{\"location\":\"Paris\"},\"name\":\"get_weather\"}<|tool_call_end|><|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			// Two assistants with tool calls - first has thinking stripped
-			name: "two assistants with tools: first thinking stripped",
-			messages: []api.Message{
-				{Role: "user", Content: "What's the weather?"},
-				{
-					Role:    "assistant",
-					Content: "<think>checking</think>",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name: "get_weather",
-								Arguments: testArgs(map[string]any{
-									"location": "Paris",
-								}),
-							},
-						},
-					},
-				},
-				{Role: "tool", Content: "22C"},
-				{Role: "assistant", Content: "<think>got result</think>It's 22C!"},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nWhat's the weather?<|im_end|>\n<|im_start|>assistant\n<|tool_call_start|>{\"arguments\":{\"location\":\"Paris\"},\"name\":\"get_weather\"}<|tool_call_end|><|im_end|>\n<|im_start|>tool\n22C<|im_end|>\n<|im_start|>assistant\n<think>got result</think>It's 22C!<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			// Two assistants with tools - both preserved when thinking enabled
-			name: "two assistants with tools: both preserved when thinking enabled",
-			messages: []api.Message{
-				{Role: "user", Content: "What's the weather?"},
-				{
-					Role:    "assistant",
-					Content: "<think>checking</think>",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name: "get_weather",
-								Arguments: testArgs(map[string]any{
-									"location": "Paris",
-								}),
-							},
-						},
-					},
-				},
-				{Role: "tool", Content: "22C"},
-				{Role: "assistant", Content: "<think>got result</think>It's 22C!"},
-			},
-			thinkValue: &api.ThinkValue{Value: true},
-			expected:   "<|im_start|>user\nWhat's the weather?<|im_end|>\n<|im_start|>assistant\n<think>checking</think><|tool_call_start|>{\"arguments\":{\"location\":\"Paris\"},\"name\":\"get_weather\"}<|tool_call_end|><|im_end|>\n<|im_start|>tool\n22C<|im_end|>\n<|im_start|>assistant\n<think>got result</think>It's 22C!<|im_end|>\n<|im_start|>assistant\n",
-		},
-		{
-			// Content before thinking before tool call
-			name: "content then thinking then tool call",
-			messages: []api.Message{
-				{Role: "user", Content: "What's the weather?"},
-				{
-					Role:    "assistant",
-					Content: "Let me check.<think>Using weather API</think>",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name: "get_weather",
-								Arguments: testArgs(map[string]any{
-									"location": "Paris",
-								}),
-							},
-						},
-					},
-				},
-			},
-			thinkValue: &api.ThinkValue{Value: false},
-			expected:   "<|im_start|>user\nWhat's the weather?<|im_end|>\n<|im_start|>assistant\nLet me check.<think>Using weather API</think><|tool_call_start|>{\"arguments\":{\"location\":\"Paris\"},\"name\":\"get_weather\"}<|tool_call_end|><|im_end|>\n<|im_start|>assistant\n",
-		},
-	}
-
-	renderer := &LFM2Renderer{IsThinking: true}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			rendered, err := renderer.Render(tt.messages, tt.tools, tt.thinkValue)
-			if err != nil {
-				t.Fatalf("Render() error = %v", err)
-			}
-			if diff := cmp.Diff(tt.expected, rendered); diff != "" {
-				t.Errorf("Render() mismatch (-want +got):\n%s", diff)
-			}
-		})
-	}
-}

model/renderers/renderer.go

@@ -80,12 +80,6 @@ func rendererForName(name string) Renderer {
 		return &Nemotron3NanoRenderer{}
 	case "functiongemma":
 		return &FunctionGemmaRenderer{}
-	case "glm-4.7":
-		return &GLM47Renderer{}
-	case "lfm2":
-		return &LFM2Renderer{IsThinking: false}
-	case "lfm2-thinking":
-		return &LFM2Renderer{IsThinking: true}
 	default:
 		return nil
 	}

model/renderers/testhelpers_test.go

@@ -1,26 +1,6 @@
 package renderers
 
-import (
-	"encoding/json"
-
-	"github.com/ollama/ollama/api"
-)
-
-func args(s string) api.ToolCallFunctionArguments {
-	var result api.ToolCallFunctionArguments
-	if err := json.Unmarshal([]byte(s), &result); err != nil {
-		panic("invalid JSON in args(): " + err.Error())
-	}
-	return result
-}
-
-func propsMap(s string) *api.ToolPropertiesMap {
-	var result api.ToolPropertiesMap
-	if err := json.Unmarshal([]byte(s), &result); err != nil {
-		panic("invalid JSON in propsMap(): " + err.Error())
-	}
-	return &result
-}
+import "github.com/ollama/ollama/api"
 
 // testPropsMap creates a ToolPropertiesMap from a map (convenience function for tests, order not preserved)
 func testPropsMap(m map[string]api.ToolProperty) *api.ToolPropertiesMap {

openai/openai.go

@@ -630,10 +630,6 @@ func nameFromToolCallID(messages []Message, toolCallID string) string {
 
 // decodeImageURL decodes a base64 data URI into raw image bytes.
 func decodeImageURL(url string) (api.ImageData, error) {
-	if strings.HasPrefix(url, "http://") || strings.HasPrefix(url, "https://") {
-		return nil, errors.New("image URLs are not currently supported, please use base64 encoded data instead")
-	}
-
 	types := []string{"jpeg", "jpg", "png", "webp"}
 
 	// Support blank mime type to match /api/chat's behavior of taking just unadorned base64
@@ -737,104 +733,3 @@ func FromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 		DebugRenderOnly: r.DebugRenderOnly,
 	}, nil
 }
-
-// ImageGenerationRequest is an OpenAI-compatible image generation request.
-type ImageGenerationRequest struct {
-	Model          string `json:"model"`
-	Prompt         string `json:"prompt"`
-	N              int    `json:"n,omitempty"`
-	Size           string `json:"size,omitempty"`
-	ResponseFormat string `json:"response_format,omitempty"`
-	Seed           *int64 `json:"seed,omitempty"`
-}
-
-// ImageGenerationResponse is an OpenAI-compatible image generation response.
-type ImageGenerationResponse struct {
-	Created int64            `json:"created"`
-	Data    []ImageURLOrData `json:"data"`
-}
-
-// ImageURLOrData contains either a URL or base64-encoded image data.
-type ImageURLOrData struct {
-	URL     string `json:"url,omitempty"`
-	B64JSON string `json:"b64_json,omitempty"`
-}
-
-// FromImageGenerationRequest converts an OpenAI image generation request to an Ollama GenerateRequest.
-func FromImageGenerationRequest(r ImageGenerationRequest) api.GenerateRequest {
-	req := api.GenerateRequest{
-		Model:  r.Model,
-		Prompt: r.Prompt,
-	}
-	// Parse size if provided (e.g., "1024x768")
-	if r.Size != "" {
-		var w, h int32
-		if _, err := fmt.Sscanf(r.Size, "%dx%d", &w, &h); err == nil {
-			req.Width = w
-			req.Height = h
-		}
-	}
-	if r.Seed != nil {
-		if req.Options == nil {
-			req.Options = map[string]any{}
-		}
-		req.Options["seed"] = *r.Seed
-	}
-	return req
-}
-
-// ToImageGenerationResponse converts an Ollama GenerateResponse to an OpenAI ImageGenerationResponse.
-func ToImageGenerationResponse(resp api.GenerateResponse) ImageGenerationResponse {
-	var data []ImageURLOrData
-	if resp.Image != "" {
-		data = []ImageURLOrData{{B64JSON: resp.Image}}
-	}
-	return ImageGenerationResponse{
-		Created: resp.CreatedAt.Unix(),
-		Data:    data,
-	}
-}
-
-// ImageEditRequest is an OpenAI-compatible image edit request.
-type ImageEditRequest struct {
-	Model  string `json:"model"`
-	Prompt string `json:"prompt"`
-	Image  string `json:"image"`          // Base64-encoded image data
-	Size   string `json:"size,omitempty"` // e.g., "1024x1024"
-	Seed   *int64 `json:"seed,omitempty"`
-}
-
-// FromImageEditRequest converts an OpenAI image edit request to an Ollama GenerateRequest.
-func FromImageEditRequest(r ImageEditRequest) (api.GenerateRequest, error) {
-	req := api.GenerateRequest{
-		Model:  r.Model,
-		Prompt: r.Prompt,
-	}
-
-	// Decode the input image
-	if r.Image != "" {
-		imgData, err := decodeImageURL(r.Image)
-		if err != nil {
-			return api.GenerateRequest{}, fmt.Errorf("invalid image: %w", err)
-		}
-		req.Images = append(req.Images, imgData)
-	}
-
-	// Parse size if provided (e.g., "1024x768")
-	if r.Size != "" {
-		var w, h int32
-		if _, err := fmt.Sscanf(r.Size, "%dx%d", &w, &h); err == nil {
-			req.Width = w
-			req.Height = h
-		}
-	}
-
-	if r.Seed != nil {
-		if req.Options == nil {
-			req.Options = map[string]any{}
-		}
-		req.Options["seed"] = *r.Seed
-	}
-
-	return req, nil
-}

openai/openai_test.go

@@ -448,86 +448,3 @@ func TestFromChatRequest_TopLogprobsRange(t *testing.T) {
 		})
 	}
 }
-
-func TestFromImageEditRequest_Basic(t *testing.T) {
-	req := ImageEditRequest{
-		Model:  "test-model",
-		Prompt: "make it blue",
-		Image:  prefix + image,
-	}
-
-	result, err := FromImageEditRequest(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 result.Prompt != "make it blue" {
-		t.Errorf("expected prompt 'make it blue', got %q", result.Prompt)
-	}
-
-	if len(result.Images) != 1 {
-		t.Fatalf("expected 1 image, got %d", len(result.Images))
-	}
-}
-
-func TestFromImageEditRequest_WithSize(t *testing.T) {
-	req := ImageEditRequest{
-		Model:  "test-model",
-		Prompt: "make it blue",
-		Image:  prefix + image,
-		Size:   "512x768",
-	}
-
-	result, err := FromImageEditRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if result.Width != 512 {
-		t.Errorf("expected width 512, got %d", result.Width)
-	}
-
-	if result.Height != 768 {
-		t.Errorf("expected height 768, got %d", result.Height)
-	}
-}
-
-func TestFromImageEditRequest_WithSeed(t *testing.T) {
-	seed := int64(12345)
-	req := ImageEditRequest{
-		Model:  "test-model",
-		Prompt: "make it blue",
-		Image:  prefix + image,
-		Seed:   &seed,
-	}
-
-	result, err := FromImageEditRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if result.Options == nil {
-		t.Fatal("expected options to be set")
-	}
-
-	if result.Options["seed"] != seed {
-		t.Errorf("expected seed %d, got %v", seed, result.Options["seed"])
-	}
-}
-
-func TestFromImageEditRequest_InvalidImage(t *testing.T) {
-	req := ImageEditRequest{
-		Model:  "test-model",
-		Prompt: "make it blue",
-		Image:  "not-valid-base64",
-	}
-
-	_, err := FromImageEditRequest(req)
-	if err == nil {
-		t.Error("expected error for invalid image")
-	}
-}

openai/responses.go

@@ -4,7 +4,6 @@ import (
 	"encoding/json"
 	"fmt"
 	"math/rand"
-	"time"
 
 	"github.com/ollama/ollama/api"
 )
@@ -266,9 +265,9 @@ type ResponsesText struct {
 type ResponsesTool struct {
 	Type        string         `json:"type"` // "function"
 	Name        string         `json:"name"`
-	Description *string        `json:"description"` // nullable but required
-	Strict      *bool          `json:"strict"`      // nullable but required
-	Parameters  map[string]any `json:"parameters"`  // nullable but required
+	Description string         `json:"description,omitempty"`
+	Strict      bool           `json:"strict,omitempty"`
+	Parameters  map[string]any `json:"parameters,omitempty"`
 }
 
 type ResponsesRequest struct {
@@ -476,16 +475,11 @@ func convertTool(t ResponsesTool) (api.Tool, error) {
 		}
 	}
 
-	var description string
-	if t.Description != nil {
-		description = *t.Description
-	}
-
 	return api.Tool{
 		Type: t.Type,
 		Function: api.ToolFunction{
 			Name:        t.Name,
-			Description: description,
+			Description: t.Description,
 			Parameters:  params,
 		},
 	}, nil
@@ -522,60 +516,17 @@ func convertInputMessage(m ResponsesInputMessage) (api.Message, error) {
 
 // Response types for the Responses API
 
-// ResponsesTextField represents the text output configuration in the response.
-type ResponsesTextField struct {
-	Format ResponsesTextFormat `json:"format"`
-}
-
-// ResponsesReasoningOutput represents reasoning configuration in the response.
-type ResponsesReasoningOutput struct {
-	Effort  *string `json:"effort,omitempty"`
-	Summary *string `json:"summary,omitempty"`
-}
-
-// ResponsesError represents an error in the response.
-type ResponsesError struct {
-	Code    string `json:"code"`
-	Message string `json:"message"`
-}
-
-// ResponsesIncompleteDetails represents details about why a response was incomplete.
-type ResponsesIncompleteDetails struct {
-	Reason string `json:"reason"`
-}
-
 type ResponsesResponse struct {
-	ID                 string                      `json:"id"`
-	Object             string                      `json:"object"`
-	CreatedAt          int64                       `json:"created_at"`
-	CompletedAt        *int64                      `json:"completed_at"`
-	Status             string                      `json:"status"`
-	IncompleteDetails  *ResponsesIncompleteDetails `json:"incomplete_details"`
-	Model              string                      `json:"model"`
-	PreviousResponseID *string                     `json:"previous_response_id"`
-	Instructions       *string                     `json:"instructions"`
-	Output             []ResponsesOutputItem       `json:"output"`
-	Error              *ResponsesError             `json:"error"`
-	Tools              []ResponsesTool             `json:"tools"`
-	ToolChoice         any                         `json:"tool_choice"`
-	Truncation         string                      `json:"truncation"`
-	ParallelToolCalls  bool                        `json:"parallel_tool_calls"`
-	Text               ResponsesTextField          `json:"text"`
-	TopP               float64                     `json:"top_p"`
-	PresencePenalty    float64                     `json:"presence_penalty"`
-	FrequencyPenalty   float64                     `json:"frequency_penalty"`
-	TopLogprobs        int                         `json:"top_logprobs"`
-	Temperature        float64                     `json:"temperature"`
-	Reasoning          *ResponsesReasoningOutput   `json:"reasoning"`
-	Usage              *ResponsesUsage             `json:"usage"`
-	MaxOutputTokens    *int                        `json:"max_output_tokens"`
-	MaxToolCalls       *int                        `json:"max_tool_calls"`
-	Store              bool                        `json:"store"`
-	Background         bool                        `json:"background"`
-	ServiceTier        string                      `json:"service_tier"`
-	Metadata           map[string]any              `json:"metadata"`
-	SafetyIdentifier   *string                     `json:"safety_identifier"`
-	PromptCacheKey     *string                     `json:"prompt_cache_key"`
+	ID        string                `json:"id"`
+	Object    string                `json:"object"`
+	CreatedAt int64                 `json:"created_at"`
+	Status    string                `json:"status"`
+	Model     string                `json:"model"`
+	Output    []ResponsesOutputItem `json:"output"`
+	Usage     *ResponsesUsage       `json:"usage,omitempty"`
+	// TODO(drifkin): add `temperature` and `top_p` to the response, but this
+	// requires additional plumbing to find the effective values since the
+	// defaults can come from the model or the request
 }
 
 type ResponsesOutputItem struct {
@@ -599,39 +550,18 @@ type ResponsesReasoningSummary struct {
 }
 
 type ResponsesOutputContent struct {
-	Type        string `json:"type"` // "output_text"
-	Text        string `json:"text"`
-	Annotations []any  `json:"annotations"`
-	Logprobs    []any  `json:"logprobs"`
-}
-
-type ResponsesInputTokensDetails struct {
-	CachedTokens int `json:"cached_tokens"`
-}
-
-type ResponsesOutputTokensDetails struct {
-	ReasoningTokens int `json:"reasoning_tokens"`
+	Type string `json:"type"` // "output_text"
+	Text string `json:"text"`
 }
 
 type ResponsesUsage struct {
-	InputTokens         int                          `json:"input_tokens"`
-	OutputTokens        int                          `json:"output_tokens"`
-	TotalTokens         int                          `json:"total_tokens"`
-	InputTokensDetails  ResponsesInputTokensDetails  `json:"input_tokens_details"`
-	OutputTokensDetails ResponsesOutputTokensDetails `json:"output_tokens_details"`
+	InputTokens  int `json:"input_tokens"`
+	OutputTokens int `json:"output_tokens"`
+	TotalTokens  int `json:"total_tokens"`
 }
 
-// derefFloat64 returns the value of a float64 pointer, or a default if nil.
-func derefFloat64(p *float64, def float64) float64 {
-	if p != nil {
-		return *p
-	}
-	return def
-}
-
-// ToResponse converts an api.ChatResponse to a Responses API response.
-// The request is used to echo back request parameters in the response.
-func ToResponse(model, responseID, itemID string, chatResponse api.ChatResponse, request ResponsesRequest) ResponsesResponse {
+// ToResponse converts an api.ChatResponse to a Responses API response
+func ToResponse(model, responseID, itemID string, chatResponse api.ChatResponse) ResponsesResponse {
 	var output []ResponsesOutputItem
 
 	// Add reasoning item if thinking is present
@@ -655,7 +585,6 @@ func ToResponse(model, responseID, itemID string, chatResponse api.ChatResponse,
 			output = append(output, ResponsesOutputItem{
 				ID:        fmt.Sprintf("fc_%s_%d", responseID, i),
 				Type:      "function_call",
-				Status:    "completed",
 				CallID:    tc.ID,
 				Name:      tc.Function.Name,
 				Arguments: tc.Function.Arguments,
@@ -669,90 +598,25 @@ func ToResponse(model, responseID, itemID string, chatResponse api.ChatResponse,
 			Role:   "assistant",
 			Content: []ResponsesOutputContent{
 				{
-					Type:        "output_text",
-					Text:        chatResponse.Message.Content,
-					Annotations: []any{},
-					Logprobs:    []any{},
+					Type: "output_text",
+					Text: chatResponse.Message.Content,
 				},
 			},
 		})
 	}
 
-	var instructions *string
-	if request.Instructions != "" {
-		instructions = &request.Instructions
-	}
-
-	// Build truncation with default
-	truncation := "disabled"
-	if request.Truncation != nil {
-		truncation = *request.Truncation
-	}
-
-	tools := request.Tools
-	if tools == nil {
-		tools = []ResponsesTool{}
-	}
-
-	text := ResponsesTextField{
-		Format: ResponsesTextFormat{Type: "text"},
-	}
-	if request.Text != nil && request.Text.Format != nil {
-		text.Format = *request.Text.Format
-	}
-
-	// Build reasoning output from request
-	var reasoning *ResponsesReasoningOutput
-	if request.Reasoning.Effort != "" || request.Reasoning.Summary != "" {
-		reasoning = &ResponsesReasoningOutput{}
-		if request.Reasoning.Effort != "" {
-			reasoning.Effort = &request.Reasoning.Effort
-		}
-		if request.Reasoning.Summary != "" {
-			reasoning.Summary = &request.Reasoning.Summary
-		}
-	}
-
 	return ResponsesResponse{
-		ID:                 responseID,
-		Object:             "response",
-		CreatedAt:          chatResponse.CreatedAt.Unix(),
-		CompletedAt:        nil, // Set by middleware when writing final response
-		Status:             "completed",
-		IncompleteDetails:  nil, // Only populated if response incomplete
-		Model:              model,
-		PreviousResponseID: nil, // Not supported
-		Instructions:       instructions,
-		Output:             output,
-		Error:              nil, // Only populated on failure
-		Tools:              tools,
-		ToolChoice:         "auto", // Default value
-		Truncation:         truncation,
-		ParallelToolCalls:  true, // Default value
-		Text:               text,
-		TopP:               derefFloat64(request.TopP, 1.0),
-		PresencePenalty:    0, // Default value
-		FrequencyPenalty:   0, // Default value
-		TopLogprobs:        0, // Default value
-		Temperature:        derefFloat64(request.Temperature, 1.0),
-		Reasoning:          reasoning,
+		ID:        responseID,
+		Object:    "response",
+		CreatedAt: chatResponse.CreatedAt.Unix(),
+		Status:    "completed",
+		Model:     model,
+		Output:    output,
 		Usage: &ResponsesUsage{
 			InputTokens:  chatResponse.PromptEvalCount,
 			OutputTokens: chatResponse.EvalCount,
 			TotalTokens:  chatResponse.PromptEvalCount + chatResponse.EvalCount,
-			// TODO(drifkin): wire through the actual values
-			InputTokensDetails: ResponsesInputTokensDetails{CachedTokens: 0},
-			// TODO(drifkin): wire through the actual values
-			OutputTokensDetails: ResponsesOutputTokensDetails{ReasoningTokens: 0},
 		},
-		MaxOutputTokens:  request.MaxOutputTokens,
-		MaxToolCalls:     nil,   // Not supported
-		Store:            false, // We don't store responses
-		Background:       request.Background,
-		ServiceTier:      "default", // Default value
-		Metadata:         map[string]any{},
-		SafetyIdentifier: nil, // Not supported
-		PromptCacheKey:   nil, // Not supported
 	}
 }
 
@@ -772,7 +636,6 @@ type ResponsesStreamConverter struct {
 	responseID string
 	itemID     string
 	model      string
-	request    ResponsesRequest
 
 	// State tracking (mutated across Process calls)
 	firstWrite      bool
@@ -805,12 +668,11 @@ func (c *ResponsesStreamConverter) newEvent(eventType string, data map[string]an
 }
 
 // NewResponsesStreamConverter creates a new converter with the given configuration.
-func NewResponsesStreamConverter(responseID, itemID, model string, request ResponsesRequest) *ResponsesStreamConverter {
+func NewResponsesStreamConverter(responseID, itemID, model string) *ResponsesStreamConverter {
 	return &ResponsesStreamConverter{
 		responseID: responseID,
 		itemID:     itemID,
 		model:      model,
-		request:    request,
 		firstWrite: true,
 	}
 }
@@ -855,120 +717,25 @@ func (c *ResponsesStreamConverter) Process(r api.ChatResponse) []ResponsesStream
 	return events
 }
 
-// buildResponseObject creates a full response object with all required fields for streaming events.
-func (c *ResponsesStreamConverter) buildResponseObject(status string, output []any, usage map[string]any) map[string]any {
-	var instructions any = nil
-	if c.request.Instructions != "" {
-		instructions = c.request.Instructions
-	}
-
-	truncation := "disabled"
-	if c.request.Truncation != nil {
-		truncation = *c.request.Truncation
-	}
-
-	var tools []any
-	if c.request.Tools != nil {
-		for _, t := range c.request.Tools {
-			tools = append(tools, map[string]any{
-				"type":        t.Type,
-				"name":        t.Name,
-				"description": t.Description,
-				"strict":      t.Strict,
-				"parameters":  t.Parameters,
-			})
-		}
-	}
-	if tools == nil {
-		tools = []any{}
-	}
-
-	textFormat := map[string]any{"type": "text"}
-	if c.request.Text != nil && c.request.Text.Format != nil {
-		textFormat = map[string]any{
-			"type": c.request.Text.Format.Type,
-		}
-		if c.request.Text.Format.Name != "" {
-			textFormat["name"] = c.request.Text.Format.Name
-		}
-		if c.request.Text.Format.Schema != nil {
-			textFormat["schema"] = c.request.Text.Format.Schema
-		}
-		if c.request.Text.Format.Strict != nil {
-			textFormat["strict"] = *c.request.Text.Format.Strict
-		}
-	}
-
-	var reasoning any = nil
-	if c.request.Reasoning.Effort != "" || c.request.Reasoning.Summary != "" {
-		r := map[string]any{}
-		if c.request.Reasoning.Effort != "" {
-			r["effort"] = c.request.Reasoning.Effort
-		} else {
-			r["effort"] = nil
-		}
-		if c.request.Reasoning.Summary != "" {
-			r["summary"] = c.request.Reasoning.Summary
-		} else {
-			r["summary"] = nil
-		}
-		reasoning = r
-	}
-
-	// Build top_p and temperature with defaults
-	topP := 1.0
-	if c.request.TopP != nil {
-		topP = *c.request.TopP
-	}
-	temperature := 1.0
-	if c.request.Temperature != nil {
-		temperature = *c.request.Temperature
-	}
-
-	return map[string]any{
-		"id":                   c.responseID,
-		"object":               "response",
-		"created_at":           time.Now().Unix(),
-		"completed_at":         nil,
-		"status":               status,
-		"incomplete_details":   nil,
-		"model":                c.model,
-		"previous_response_id": nil,
-		"instructions":         instructions,
-		"output":               output,
-		"error":                nil,
-		"tools":                tools,
-		"tool_choice":          "auto",
-		"truncation":           truncation,
-		"parallel_tool_calls":  true,
-		"text":                 map[string]any{"format": textFormat},
-		"top_p":                topP,
-		"presence_penalty":     0,
-		"frequency_penalty":    0,
-		"top_logprobs":         0,
-		"temperature":          temperature,
-		"reasoning":            reasoning,
-		"usage":                usage,
-		"max_output_tokens":    c.request.MaxOutputTokens,
-		"max_tool_calls":       nil,
-		"store":                false,
-		"background":           c.request.Background,
-		"service_tier":         "default",
-		"metadata":             map[string]any{},
-		"safety_identifier":    nil,
-		"prompt_cache_key":     nil,
-	}
-}
-
 func (c *ResponsesStreamConverter) createResponseCreatedEvent() ResponsesStreamEvent {
 	return c.newEvent("response.created", map[string]any{
-		"response": c.buildResponseObject("in_progress", []any{}, nil),
+		"response": map[string]any{
+			"id":     c.responseID,
+			"object": "response",
+			"status": "in_progress",
+			"output": []any{},
+		},
 	})
 }
 
 func (c *ResponsesStreamConverter) createResponseInProgressEvent() ResponsesStreamEvent {
 	return c.newEvent("response.in_progress", map[string]any{
-		"response": c.buildResponseObject("in_progress", []any{}, nil),
+		"response": map[string]any{
+			"id":     c.responseID,
+			"object": "response",
+			"status": "in_progress",
+			"output": []any{},
+		},
 	})
 }
 
@@ -995,10 +762,9 @@ func (c *ResponsesStreamConverter) processThinking(thinking string) []ResponsesS
 
 	// Emit delta
 	events = append(events, c.newEvent("response.reasoning_summary_text.delta", map[string]any{
-		"item_id":       c.reasoningItemID,
-		"output_index":  c.outputIndex,
-		"summary_index": 0,
-		"delta":         thinking,
+		"item_id":      c.reasoningItemID,
+		"output_index": c.outputIndex,
+		"delta":        thinking,
 	}))
 
 	// TODO(drifkin): consider adding
@@ -1017,10 +783,9 @@ func (c *ResponsesStreamConverter) finishReasoning() []ResponsesStreamEvent {
 
 	events := []ResponsesStreamEvent{
 		c.newEvent("response.reasoning_summary_text.done", map[string]any{
-			"item_id":       c.reasoningItemID,
-			"output_index":  c.outputIndex,
-			"summary_index": 0,
-			"text":          c.accumulatedThinking,
+			"item_id":      c.reasoningItemID,
+			"output_index": c.outputIndex,
+			"text":         c.accumulatedThinking,
 		}),
 		c.newEvent("response.output_item.done", map[string]any{
 			"output_index": c.outputIndex,
@@ -1133,10 +898,8 @@ func (c *ResponsesStreamConverter) processTextContent(content string) []Response
 			"output_index":  c.outputIndex,
 			"content_index": c.contentIndex,
 			"part": map[string]any{
-				"type":        "output_text",
-				"text":        "",
-				"annotations": []any{},
-				"logprobs":    []any{},
+				"type": "output_text",
+				"text": "",
 			},
 		}))
 	}
@@ -1150,7 +913,6 @@ func (c *ResponsesStreamConverter) processTextContent(content string) []Response
 		"output_index":  c.outputIndex,
 		"content_index": 0,
 		"delta":         content,
-		"logprobs":      []any{},
 	}))
 
 	return events
@@ -1182,10 +944,8 @@ func (c *ResponsesStreamConverter) buildFinalOutput() []any {
 			"status": "completed",
 			"role":   "assistant",
 			"content": []map[string]any{{
-				"type":        "output_text",
-				"text":        c.accumulatedText,
-				"annotations": []any{},
-				"logprobs":    []any{},
+				"type": "output_text",
+				"text": c.accumulatedText,
 			}},
 		})
 	}
@@ -1207,7 +967,6 @@ func (c *ResponsesStreamConverter) processCompletion(r api.ChatResponse) []Respo
 			"output_index":  c.outputIndex,
 			"content_index": 0,
 			"text":          c.accumulatedText,
-			"logprobs":      []any{},
 		}))
 
 		// response.content_part.done
@@ -1216,10 +975,8 @@ func (c *ResponsesStreamConverter) processCompletion(r api.ChatResponse) []Respo
 			"output_index":  c.outputIndex,
 			"content_index": 0,
 			"part": map[string]any{
-				"type":        "output_text",
-				"text":        c.accumulatedText,
-				"annotations": []any{},
-				"logprobs":    []any{},
+				"type": "output_text",
+				"text": c.accumulatedText,
 			},
 		}))
 
@@ -1232,31 +989,26 @@ func (c *ResponsesStreamConverter) processCompletion(r api.ChatResponse) []Respo
 				"status": "completed",
 				"role":   "assistant",
 				"content": []map[string]any{{
-					"type":        "output_text",
-					"text":        c.accumulatedText,
-					"annotations": []any{},
-					"logprobs":    []any{},
+					"type": "output_text",
+					"text": c.accumulatedText,
 				}},
 			},
 		}))
 	}
 
 	// response.completed
-	usage := map[string]any{
-		"input_tokens":  r.PromptEvalCount,
-		"output_tokens": r.EvalCount,
-		"total_tokens":  r.PromptEvalCount + r.EvalCount,
-		"input_tokens_details": map[string]any{
-			"cached_tokens": 0,
-		},
-		"output_tokens_details": map[string]any{
-			"reasoning_tokens": 0,
-		},
-	}
-	response := c.buildResponseObject("completed", c.buildFinalOutput(), usage)
-	response["completed_at"] = time.Now().Unix()
 	events = append(events, c.newEvent("response.completed", map[string]any{
-		"response": response,
+		"response": map[string]any{
+			"id":     c.responseID,
+			"object": "response",
+			"status": "completed",
+			"output": c.buildFinalOutput(),
+			"usage": map[string]any{
+				"input_tokens":  r.PromptEvalCount,
+				"output_tokens": r.EvalCount,
+				"total_tokens":  r.PromptEvalCount + r.EvalCount,
+			},
+		},
 	}))
 
 	return events

openai/responses_test.go

@@ -850,7 +850,7 @@ func TestFromResponsesRequest_Images(t *testing.T) {
 }
 
 func TestResponsesStreamConverter_TextOnly(t *testing.T) {
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
 
 	// First chunk with content
 	events := converter.Process(api.ChatResponse{
@@ -916,7 +916,7 @@ func TestResponsesStreamConverter_TextOnly(t *testing.T) {
 }
 
 func TestResponsesStreamConverter_ToolCalls(t *testing.T) {
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
 
 	events := converter.Process(api.ChatResponse{
 		Message: api.Message{
@@ -952,7 +952,7 @@ func TestResponsesStreamConverter_ToolCalls(t *testing.T) {
 }
 
 func TestResponsesStreamConverter_Reasoning(t *testing.T) {
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
 
 	// First chunk with thinking
 	events := converter.Process(api.ChatResponse{
@@ -1267,7 +1267,7 @@ func TestToResponse_WithReasoning(t *testing.T) {
 			Content:  "The answer is 42",
 		},
 		Done: true,
-	}, ResponsesRequest{})
+	})
 
 	// Should have 2 output items: reasoning + message
 	if len(response.Output) != 2 {
@@ -1638,7 +1638,7 @@ func TestFromResponsesRequest_ShorthandFormats(t *testing.T) {
 
 func TestResponsesStreamConverter_OutputIncludesContent(t *testing.T) {
 	// Verify that response.output_item.done includes content field for messages
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
 
 	// First chunk
 	converter.Process(api.ChatResponse{
@@ -1686,7 +1686,7 @@ func TestResponsesStreamConverter_OutputIncludesContent(t *testing.T) {
 
 func TestResponsesStreamConverter_ResponseCompletedIncludesOutput(t *testing.T) {
 	// Verify that response.completed includes the output array
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
 
 	// Process some content
 	converter.Process(api.ChatResponse{
@@ -1730,7 +1730,7 @@ func TestResponsesStreamConverter_ResponseCompletedIncludesOutput(t *testing.T)
 
 func TestResponsesStreamConverter_ResponseCreatedIncludesOutput(t *testing.T) {
 	// Verify that response.created includes an empty output array
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
 
 	events := converter.Process(api.ChatResponse{
 		Message: api.Message{Content: "Hi"},
@@ -1757,7 +1757,7 @@ func TestResponsesStreamConverter_ResponseCreatedIncludesOutput(t *testing.T) {
 
 func TestResponsesStreamConverter_SequenceNumbers(t *testing.T) {
 	// Verify that events include incrementing sequence numbers
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
 
 	events := converter.Process(api.ChatResponse{
 		Message: api.Message{Content: "Hello"},
@@ -1791,7 +1791,7 @@ func TestResponsesStreamConverter_SequenceNumbers(t *testing.T) {
 
 func TestResponsesStreamConverter_FunctionCallStatus(t *testing.T) {
 	// Verify that function call items include status field
-	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b", ResponsesRequest{})
+	converter := NewResponsesStreamConverter("resp_123", "msg_456", "gpt-oss:20b")
 
 	events := converter.Process(api.ChatResponse{
 		Message: api.Message{

readline/readline.go

@@ -5,7 +5,6 @@ import (
 	"fmt"
 	"io"
 	"os"
-	"strings"
 )
 
 type Prompt struct {
@@ -37,11 +36,10 @@ type Terminal struct {
 }
 
 type Instance struct {
-	Prompt      *Prompt
-	Terminal    *Terminal
-	History     *History
-	Pasting     bool
-	pastedLines []string
+	Prompt   *Prompt
+	Terminal *Terminal
+	History  *History
+	Pasting  bool
 }
 
 func New(prompt Prompt) (*Instance, error) {
@@ -95,21 +93,7 @@ func (i *Instance) Readline() (string, error) {
 
 	var currentLineBuf []rune
 
-	// draining tracks if we're processing buffered input from cooked mode.
-	// In cooked mode Enter sends \n, but in raw mode Ctrl+J sends \n.
-	// We treat \n from cooked mode as submit, not multiline.
-	// We check Buffered() after the first read since the bufio buffer is
-	// empty until then. This is compatible with """ multiline mode in
-	// interactive.go since each Readline() call is independent.
-	var draining, stopDraining bool
-
 	for {
-		// Apply deferred state change from previous iteration
-		if stopDraining {
-			draining = false
-			stopDraining = false
-		}
-
 		// don't show placeholder when pasting unless we're in multiline mode
 		showPlaceholder := !i.Pasting || i.Prompt.UseAlt
 		if buf.IsEmpty() && showPlaceholder {
@@ -119,15 +103,6 @@ func (i *Instance) Readline() (string, error) {
 
 		r, err := i.Terminal.Read()
 
-		// After reading, check if there's more buffered data. If so, we're
-		// processing cooked-mode input. Once buffer empties, the current
-		// char is the last buffered one (still drain it), then stop next iteration.
-		if i.Terminal.reader.Buffered() > 0 {
-			draining = true
-		} else if draining {
-			stopDraining = true
-		}
-
 		if buf.IsEmpty() {
 			fmt.Print(ClearToEOL)
 		}
@@ -199,8 +174,6 @@ func (i *Instance) Readline() (string, error) {
 		case CharEsc:
 			esc = true
 		case CharInterrupt:
-			i.pastedLines = nil
-			i.Prompt.UseAlt = false
 			return "", ErrInterrupt
 		case CharPrev:
 			i.historyPrev(buf, &currentLineBuf)
@@ -215,23 +188,7 @@ func (i *Instance) Readline() (string, error) {
 		case CharForward:
 			buf.MoveRight()
 		case CharBackspace, CharCtrlH:
-			if buf.IsEmpty() && len(i.pastedLines) > 0 {
-				lastIdx := len(i.pastedLines) - 1
-				prevLine := i.pastedLines[lastIdx]
-				i.pastedLines = i.pastedLines[:lastIdx]
-				fmt.Print(CursorBOL + ClearToEOL + CursorUp + CursorBOL + ClearToEOL)
-				if len(i.pastedLines) == 0 {
-					fmt.Print(i.Prompt.Prompt)
-					i.Prompt.UseAlt = false
-				} else {
-					fmt.Print(i.Prompt.AltPrompt)
-				}
-				for _, r := range prevLine {
-					buf.Add(r)
-				}
-			} else {
-				buf.Remove()
-			}
+			buf.Remove()
 		case CharTab:
 			// todo: convert back to real tabs
 			for range 8 {
@@ -254,33 +211,13 @@ func (i *Instance) Readline() (string, error) {
 		case CharCtrlZ:
 			fd := os.Stdin.Fd()
 			return handleCharCtrlZ(fd, i.Terminal.termios)
-		case CharCtrlJ:
-			// If not draining cooked-mode input, treat as multiline
-			if !draining {
-				i.pastedLines = append(i.pastedLines, buf.String())
-				buf.Buf.Clear()
-				buf.Pos = 0
-				buf.DisplayPos = 0
-				buf.LineHasSpace.Clear()
-				fmt.Println()
-				fmt.Print(i.Prompt.AltPrompt)
-				i.Prompt.UseAlt = true
-				continue
-			}
-			// Draining cooked-mode input: treat \n as submit
-			fallthrough
-		case CharEnter:
+		case CharEnter, CharCtrlJ:
 			output := buf.String()
-			if len(i.pastedLines) > 0 {
-				output = strings.Join(i.pastedLines, "\n") + "\n" + output
-				i.pastedLines = nil
-			}
 			if output != "" {
 				i.History.Add(output)
 			}
 			buf.MoveToEnd()
 			fmt.Println()
-			i.Prompt.UseAlt = false
 
 			return output, nil
 		default:

scripts/build_darwin.sh

@@ -60,7 +60,7 @@ _build_darwin() {
             cmake --install $BUILD_DIR --component MLX
             # Override CGO flags to point to the amd64 build directory
             MLX_CGO_CFLAGS="-O3 -I$(pwd)/$BUILD_DIR/_deps/mlx-c-src -mmacosx-version-min=14.0"
-            MLX_CGO_LDFLAGS="-ldl -lc++ -framework Accelerate -mmacosx-version-min=14.0"
+            MLX_CGO_LDFLAGS="-L$(pwd)/$BUILD_DIR/lib/ollama -lmlxc -lmlx -Wl,-rpath,@executable_path -lc++ -framework Accelerate -mmacosx-version-min=14.0"
         else
             BUILD_DIR=build
             cmake --preset MLX \
@@ -71,12 +71,10 @@ _build_darwin() {
             cmake --install $BUILD_DIR --component MLX
             # Use default CGO flags from mlx.go for arm64
             MLX_CGO_CFLAGS="-O3 -I$(pwd)/$BUILD_DIR/_deps/mlx-c-src -mmacosx-version-min=14.0"
-            MLX_CGO_LDFLAGS="-lc++ -framework Metal -framework Foundation -framework Accelerate -mmacosx-version-min=14.0"
+            MLX_CGO_LDFLAGS="-L$(pwd)/$BUILD_DIR/lib/ollama -lmlxc -lmlx -Wl,-rpath,@executable_path -lc++ -framework Metal -framework Foundation -framework Accelerate -mmacosx-version-min=14.0"
         fi
-        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 CGO_CFLAGS="$MLX_CGO_CFLAGS" CGO_LDFLAGS="$MLX_CGO_LDFLAGS" go build -tags mlx -o $INSTALL_PREFIX .
-        # Copy MLX libraries to same directory as executable for dlopen
-        cp $INSTALL_PREFIX/lib/ollama/libmlxc.dylib $INSTALL_PREFIX/
-        cp $INSTALL_PREFIX/lib/ollama/libmlx.dylib $INSTALL_PREFIX/
+        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 CGO_CFLAGS="$MLX_CGO_CFLAGS" CGO_LDFLAGS="$MLX_CGO_LDFLAGS" go build -tags mlx -o $INSTALL_PREFIX/ollama-mlx .
+        GOOS=darwin GOARCH=$ARCH CGO_ENABLED=1 go build -o $INSTALL_PREFIX .
     done
 }
 
@@ -84,10 +82,12 @@ _sign_darwin() {
     status "Creating universal binary..."
     mkdir -p dist/darwin
     lipo -create -output dist/darwin/ollama dist/darwin-*/ollama
+    lipo -create -output dist/darwin/ollama-mlx dist/darwin-*/ollama-mlx
     chmod +x dist/darwin/ollama
+    chmod +x dist/darwin/ollama-mlx
 
     if [ -n "$APPLE_IDENTITY" ]; then
-        for F in dist/darwin/ollama dist/darwin-*/lib/ollama/*; do
+        for F in dist/darwin/ollama dist/darwin-*/lib/ollama/* dist/darwin/ollama-mlx; do
             codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime $F
         done
 
@@ -154,6 +154,7 @@ _build_macapp() {
     mkdir -p dist/Ollama.app/Contents/Resources
     if [ -d dist/darwin-amd64 ]; then
         lipo -create -output dist/Ollama.app/Contents/Resources/ollama dist/darwin-amd64/ollama dist/darwin-arm64/ollama
+        lipo -create -output dist/Ollama.app/Contents/Resources/ollama-mlx dist/darwin-amd64/ollama-mlx dist/darwin-arm64/ollama-mlx
         for F in dist/darwin-amd64/lib/ollama/*mlx*.dylib ; do
             lipo -create -output dist/darwin/$(basename $F) $F dist/darwin-arm64/lib/ollama/$(basename $F)
         done
@@ -165,27 +166,28 @@ _build_macapp() {
         cp -a dist/darwin/ollama dist/Ollama.app/Contents/Resources/ollama
         cp dist/darwin/*.so dist/darwin/*.dylib dist/Ollama.app/Contents/Resources/
     fi
+    cp -a dist/darwin/ollama-mlx dist/Ollama.app/Contents/Resources/ollama-mlx
     chmod a+x dist/Ollama.app/Contents/Resources/ollama
 
     # Sign
     if [ -n "$APPLE_IDENTITY" ]; then
         codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime dist/Ollama.app/Contents/Resources/ollama
-        for lib in dist/Ollama.app/Contents/Resources/*.so dist/Ollama.app/Contents/Resources/*.dylib dist/Ollama.app/Contents/Resources/*.metallib ; do
+        for lib in dist/Ollama.app/Contents/Resources/*.so dist/Ollama.app/Contents/Resources/*.dylib dist/Ollama.app/Contents/Resources/*.metallib dist/Ollama.app/Contents/Resources/ollama-mlx ; do
             codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier ai.ollama.ollama --options=runtime ${lib}
         done
         codesign -f --timestamp -s "$APPLE_IDENTITY" --identifier com.electron.ollama --deep --options=runtime dist/Ollama.app
     fi
 
     rm -f dist/Ollama-darwin.zip
-    ditto -c -k --norsrc --keepParent dist/Ollama.app dist/Ollama-darwin.zip
-    (cd dist/Ollama.app/Contents/Resources/; tar -cf - ollama *.so *.dylib *.metallib 2>/dev/null) | gzip -9vc > dist/ollama-darwin.tgz
+    ditto -c -k --keepParent dist/Ollama.app dist/Ollama-darwin.zip
+    (cd dist/Ollama.app/Contents/Resources/; tar -cf - ollama ollama-mlx *.so *.dylib *.metallib 2>/dev/null) | gzip -9vc > dist/ollama-darwin.tgz
 
     # Notarize and Staple
     if [ -n "$APPLE_IDENTITY" ]; then
         $(xcrun -f notarytool) submit dist/Ollama-darwin.zip --wait --timeout 20m --apple-id "$APPLE_ID" --password "$APPLE_PASSWORD" --team-id "$APPLE_TEAM_ID"
         rm -f dist/Ollama-darwin.zip
         $(xcrun -f stapler) staple dist/Ollama.app
-        ditto -c -k --norsrc --keepParent dist/Ollama.app dist/Ollama-darwin.zip
+        ditto -c -k --keepParent dist/Ollama.app dist/Ollama-darwin.zip
 
         rm -f dist/Ollama.dmg
 

server/auth.go

@@ -50,17 +50,12 @@ func (r registryChallenge) URL() (*url.URL, error) {
 	return redirectURL, nil
 }
 
-func getAuthorizationToken(ctx context.Context, challenge registryChallenge, originalHost string) (string, error) {
+func getAuthorizationToken(ctx context.Context, challenge registryChallenge) (string, error) {
 	redirectURL, err := challenge.URL()
 	if err != nil {
 		return "", err
 	}
 
-	// Validate that the realm host matches the original request host to prevent sending tokens cross-origin.
-	if redirectURL.Host != originalHost {
-		return "", fmt.Errorf("realm host %q does not match original host %q", redirectURL.Host, originalHost)
-	}
-
 	sha256sum := sha256.Sum256(nil)
 	data := []byte(fmt.Sprintf("%s,%s,%s", http.MethodGet, redirectURL.String(), base64.StdEncoding.EncodeToString([]byte(hex.EncodeToString(sha256sum[:])))))
 

server/auth_test.go

@@ -1,113 +0,0 @@
-package server
-
-import (
-	"context"
-	"strings"
-	"testing"
-	"time"
-)
-
-func TestGetAuthorizationTokenRejectsCrossDomain(t *testing.T) {
-	tests := []struct {
-		realm        string
-		originalHost string
-		wantMismatch bool
-	}{
-		{"https://example.com/token", "example.com", false},
-		{"https://example.com/token", "other.com", true},
-		{"https://example.com/token", "localhost:8000", true},
-		{"https://localhost:5000/token", "localhost:5000", false},
-		{"https://localhost:5000/token", "localhost:6000", true},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.originalHost, func(t *testing.T) {
-			ctx, cancel := context.WithTimeout(context.Background(), 100*time.Millisecond)
-			defer cancel()
-
-			challenge := registryChallenge{Realm: tt.realm, Service: "test", Scope: "repo:x:pull"}
-			_, err := getAuthorizationToken(ctx, challenge, tt.originalHost)
-
-			isMismatch := err != nil && strings.Contains(err.Error(), "does not match")
-			if tt.wantMismatch && !isMismatch {
-				t.Errorf("expected domain mismatch error, got: %v", err)
-			}
-			if !tt.wantMismatch && isMismatch {
-				t.Errorf("unexpected domain mismatch error: %v", err)
-			}
-		})
-	}
-}
-
-func TestParseRegistryChallenge(t *testing.T) {
-	tests := []struct {
-		input                             string
-		wantRealm, wantService, wantScope string
-	}{
-		{
-			`Bearer realm="https://auth.example.com/token",service="registry",scope="repo:foo:pull"`,
-			"https://auth.example.com/token", "registry", "repo:foo:pull",
-		},
-		{
-			`Bearer realm="https://r.ollama.ai/v2/token",service="ollama",scope="-"`,
-			"https://r.ollama.ai/v2/token", "ollama", "-",
-		},
-		{"", "", "", ""},
-	}
-
-	for _, tt := range tests {
-		result := parseRegistryChallenge(tt.input)
-		if result.Realm != tt.wantRealm || result.Service != tt.wantService || result.Scope != tt.wantScope {
-			t.Errorf("parseRegistryChallenge(%q) = {%q, %q, %q}, want {%q, %q, %q}",
-				tt.input, result.Realm, result.Service, result.Scope,
-				tt.wantRealm, tt.wantService, tt.wantScope)
-		}
-	}
-}
-
-func TestRegistryChallengeURL(t *testing.T) {
-	challenge := registryChallenge{
-		Realm:   "https://auth.example.com/token",
-		Service: "registry",
-		Scope:   "repo:foo:pull repo:bar:push",
-	}
-
-	u, err := challenge.URL()
-	if err != nil {
-		t.Fatalf("URL() error: %v", err)
-	}
-
-	if u.Host != "auth.example.com" {
-		t.Errorf("host = %q, want %q", u.Host, "auth.example.com")
-	}
-	if u.Path != "/token" {
-		t.Errorf("path = %q, want %q", u.Path, "/token")
-	}
-
-	q := u.Query()
-	if q.Get("service") != "registry" {
-		t.Errorf("service = %q, want %q", q.Get("service"), "registry")
-	}
-	if scopes := q["scope"]; len(scopes) != 2 {
-		t.Errorf("scope count = %d, want 2", len(scopes))
-	}
-	if q.Get("ts") == "" {
-		t.Error("missing ts")
-	}
-	if q.Get("nonce") == "" {
-		t.Error("missing nonce")
-	}
-
-	// Nonces should differ between calls
-	u2, _ := challenge.URL()
-	if q.Get("nonce") == u2.Query().Get("nonce") {
-		t.Error("nonce should be unique per call")
-	}
-}
-
-func TestRegistryChallengeURLInvalid(t *testing.T) {
-	challenge := registryChallenge{Realm: "://invalid"}
-	if _, err := challenge.URL(); err == nil {
-		t.Error("expected error for invalid URL")
-	}
-}

server/create.go

@@ -28,7 +28,6 @@ import (
 	"github.com/ollama/ollama/format"
 	ofs "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
-	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
@@ -91,7 +90,7 @@ func (s *Server) CreateHandler(c *gin.Context) {
 			ch <- resp
 		}
 
-		oldManifest, _ := manifest.ParseNamedManifest(name)
+		oldManifest, _ := ParseNamedManifest(name)
 
 		var baseLayers []*layerGGML
 		var err error
@@ -124,9 +123,9 @@ func (s *Server) CreateHandler(c *gin.Context) {
 				}
 
 				if err == nil && !remote && (config.Renderer == "" || config.Parser == "" || config.Requires == "") {
-					mf, mErr := manifest.ParseNamedManifest(fromName)
-					if mErr == nil && mf.Config.Digest != "" {
-						configPath, pErr := manifest.BlobsPath(mf.Config.Digest)
+					manifest, mErr := ParseNamedManifest(fromName)
+					if mErr == nil && manifest.Config.Digest != "" {
+						configPath, pErr := GetBlobsPath(manifest.Config.Digest)
 						if pErr == nil {
 							if cfgFile, fErr := os.Open(configPath); fErr == nil {
 								var baseConfig model.ConfigV2
@@ -343,7 +342,7 @@ func detectModelTypeFromFiles(files map[string]string) string {
 			return "gguf"
 		} else {
 			// try to see if we can find a gguf file even without the file extension
-			blobPath, err := manifest.BlobsPath(files[fn])
+			blobPath, err := GetBlobsPath(files[fn])
 			if err != nil {
 				slog.Error("error getting blobs path", "file", fn)
 				return ""
@@ -395,7 +394,7 @@ func convertFromSafetensors(files map[string]string, baseLayers []*layerGGML, is
 			return nil, fmt.Errorf("%w: %s: %s", errFilePath, err, fp)
 		}
 
-		blobPath, err := manifest.BlobsPath(digest)
+		blobPath, err := GetBlobsPath(digest)
 		if err != nil {
 			return nil, err
 		}
@@ -433,7 +432,7 @@ func convertFromSafetensors(files map[string]string, baseLayers []*layerGGML, is
 		return nil, err
 	}
 
-	layer, err := manifest.NewLayer(t, mediaType)
+	layer, err := NewLayer(t, mediaType)
 	if err != nil {
 		return nil, err
 	}
@@ -466,7 +465,7 @@ func kvFromLayers(baseLayers []*layerGGML) (ofs.Config, error) {
 }
 
 func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML, config *model.ConfigV2, fn func(resp api.ProgressResponse)) (err error) {
-	var layers []manifest.Layer
+	var layers []Layer
 	for _, layer := range baseLayers {
 		if layer.GGML != nil {
 			quantType := strings.ToUpper(cmp.Or(r.Quantize, r.Quantization))
@@ -551,13 +550,13 @@ func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML,
 	}
 
 	for _, layer := range layers {
-		if layer.Status != "" {
-			fn(api.ProgressResponse{Status: layer.Status})
+		if layer.status != "" {
+			fn(api.ProgressResponse{Status: layer.status})
 		}
 	}
 
 	fn(api.ProgressResponse{Status: "writing manifest"})
-	if err := manifest.WriteManifest(name, *configLayer, layers); err != nil {
+	if err := WriteManifest(name, *configLayer, layers); err != nil {
 		return err
 	}
 
@@ -578,7 +577,7 @@ func quantizeLayer(layer *layerGGML, quantizeType string, fn func(resp api.Progr
 		return nil, err
 	}
 
-	blob, err := manifest.BlobsPath(layer.Digest)
+	blob, err := GetBlobsPath(layer.Digest)
 	if err != nil {
 		return nil, err
 	}
@@ -600,7 +599,7 @@ func quantizeLayer(layer *layerGGML, quantizeType string, fn func(resp api.Progr
 	}
 	temp.Seek(0, io.SeekStart)
 	fn(api.ProgressResponse{Status: "verifying conversion"})
-	newLayer, err := manifest.NewLayer(temp, layer.MediaType)
+	newLayer, err := NewLayer(temp, layer.MediaType)
 	if err != nil {
 		return nil, err
 	}
@@ -620,7 +619,7 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 	var layers []*layerGGML
 
 	fn(api.ProgressResponse{Status: "parsing GGUF"})
-	blobPath, err := manifest.BlobsPath(digest)
+	blobPath, err := GetBlobsPath(digest)
 	if err != nil {
 		return nil, err
 	}
@@ -655,7 +654,7 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 		mediatype = "application/vnd.ollama.image.projector"
 	}
 
-	layer, err := manifest.NewLayerFromLayer(digest, mediatype, blob.Name())
+	layer, err := NewLayerFromLayer(digest, mediatype, blob.Name())
 	if err != nil {
 		slog.Debug("could not create new layer from layer", "error", err)
 		return nil, err
@@ -666,8 +665,8 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 	return detectChatTemplate(layers)
 }
 
-func removeLayer(layers []manifest.Layer, mediatype string) []manifest.Layer {
-	return slices.DeleteFunc(layers, func(layer manifest.Layer) bool {
+func removeLayer(layers []Layer, mediatype string) []Layer {
+	return slices.DeleteFunc(layers, func(layer Layer) bool {
 		if layer.MediaType != mediatype {
 			return false
 		}
@@ -681,7 +680,7 @@ func removeLayer(layers []manifest.Layer, mediatype string) []manifest.Layer {
 	})
 }
 
-func setTemplate(layers []manifest.Layer, t string) ([]manifest.Layer, error) {
+func setTemplate(layers []Layer, t string) ([]Layer, error) {
 	layers = removeLayer(layers, "application/vnd.ollama.image.template")
 	if _, err := template.Parse(t); err != nil {
 		return nil, fmt.Errorf("%w: %s", errBadTemplate, err)
@@ -691,7 +690,7 @@ func setTemplate(layers []manifest.Layer, t string) ([]manifest.Layer, error) {
 	}
 
 	blob := strings.NewReader(t)
-	layer, err := manifest.NewLayer(blob, "application/vnd.ollama.image.template")
+	layer, err := NewLayer(blob, "application/vnd.ollama.image.template")
 	if err != nil {
 		return nil, err
 	}
@@ -700,11 +699,11 @@ func setTemplate(layers []manifest.Layer, t string) ([]manifest.Layer, error) {
 	return layers, nil
 }
 
-func setSystem(layers []manifest.Layer, s string) ([]manifest.Layer, error) {
+func setSystem(layers []Layer, s string) ([]Layer, error) {
 	layers = removeLayer(layers, "application/vnd.ollama.image.system")
 	if s != "" {
 		blob := strings.NewReader(s)
-		layer, err := manifest.NewLayer(blob, "application/vnd.ollama.image.system")
+		layer, err := NewLayer(blob, "application/vnd.ollama.image.system")
 		if err != nil {
 			return nil, err
 		}
@@ -713,9 +712,9 @@ func setSystem(layers []manifest.Layer, s string) ([]manifest.Layer, error) {
 	return layers, nil
 }
 
-func setLicense(layers []manifest.Layer, l string) ([]manifest.Layer, error) {
+func setLicense(layers []Layer, l string) ([]Layer, error) {
 	blob := strings.NewReader(l)
-	layer, err := manifest.NewLayer(blob, "application/vnd.ollama.image.license")
+	layer, err := NewLayer(blob, "application/vnd.ollama.image.license")
 	if err != nil {
 		return nil, err
 	}
@@ -723,7 +722,7 @@ func setLicense(layers []manifest.Layer, l string) ([]manifest.Layer, error) {
 	return layers, nil
 }
 
-func setParameters(layers []manifest.Layer, p map[string]any) ([]manifest.Layer, error) {
+func setParameters(layers []Layer, p map[string]any) ([]Layer, error) {
 	if p == nil {
 		p = make(map[string]any)
 	}
@@ -732,7 +731,7 @@ func setParameters(layers []manifest.Layer, p map[string]any) ([]manifest.Layer,
 			continue
 		}
 
-		digestPath, err := manifest.BlobsPath(layer.Digest)
+		digestPath, err := GetBlobsPath(layer.Digest)
 		if err != nil {
 			return nil, err
 		}
@@ -766,7 +765,7 @@ func setParameters(layers []manifest.Layer, p map[string]any) ([]manifest.Layer,
 	if err := json.NewEncoder(&b).Encode(p); err != nil {
 		return nil, err
 	}
-	layer, err := manifest.NewLayer(&b, "application/vnd.ollama.image.params")
+	layer, err := NewLayer(&b, "application/vnd.ollama.image.params")
 	if err != nil {
 		return nil, err
 	}
@@ -774,7 +773,7 @@ func setParameters(layers []manifest.Layer, p map[string]any) ([]manifest.Layer,
 	return layers, nil
 }
 
-func setMessages(layers []manifest.Layer, m []api.Message) ([]manifest.Layer, error) {
+func setMessages(layers []Layer, m []api.Message) ([]Layer, error) {
 	// this leaves the old messages intact if no new messages were specified
 	// which may not be the correct behaviour
 	if len(m) == 0 {
@@ -787,7 +786,7 @@ func setMessages(layers []manifest.Layer, m []api.Message) ([]manifest.Layer, er
 	if err := json.NewEncoder(&b).Encode(m); err != nil {
 		return nil, err
 	}
-	layer, err := manifest.NewLayer(&b, "application/vnd.ollama.image.messages")
+	layer, err := NewLayer(&b, "application/vnd.ollama.image.messages")
 	if err != nil {
 		return nil, err
 	}
@@ -795,7 +794,7 @@ func setMessages(layers []manifest.Layer, m []api.Message) ([]manifest.Layer, er
 	return layers, nil
 }
 
-func createConfigLayer(layers []manifest.Layer, config model.ConfigV2) (*manifest.Layer, error) {
+func createConfigLayer(layers []Layer, config model.ConfigV2) (*Layer, error) {
 	digests := make([]string, len(layers))
 	for i, layer := range layers {
 		digests[i] = layer.Digest
@@ -806,7 +805,7 @@ func createConfigLayer(layers []manifest.Layer, config model.ConfigV2) (*manifes
 	if err := json.NewEncoder(&b).Encode(config); err != nil {
 		return nil, err
 	}
-	layer, err := manifest.NewLayer(&b, "application/vnd.docker.container.image.v1+json")
+	layer, err := NewLayer(&b, "application/vnd.docker.container.image.v1+json")
 	if err != nil {
 		return nil, err
 	}

server/create_test.go

@@ -10,7 +10,6 @@ import (
 	"testing"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/manifest"
 )
 
 func TestConvertFromSafetensors(t *testing.T) {
@@ -18,7 +17,7 @@ func TestConvertFromSafetensors(t *testing.T) {
 
 	// Helper function to create a new layer and return its digest
 	makeTemp := func(content string) string {
-		l, err := manifest.NewLayer(strings.NewReader(content), "application/octet-stream")
+		l, err := NewLayer(strings.NewReader(content), "application/octet-stream")
 		if err != nil {
 			t.Fatalf("Failed to create layer: %v", err)
 		}

server/download.go

@@ -24,8 +24,6 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
-	"github.com/ollama/ollama/manifest"
-	"github.com/ollama/ollama/types/model"
 )
 
 const maxRetries = 6
@@ -97,11 +95,48 @@ func (p *blobDownloadPart) UnmarshalJSON(b []byte) error {
 }
 
 const (
-	numDownloadParts          = 16
+	// numDownloadParts is the default number of concurrent download parts for standard downloads
+	numDownloadParts = 16
+	// numHFDownloadParts is the reduced number of concurrent download parts for HuggingFace
+	// downloads to avoid triggering rate limits (HTTP 429 errors). See GitHub issue #13297.
+	numHFDownloadParts        = 4
 	minDownloadPartSize int64 = 100 * format.MegaByte
 	maxDownloadPartSize int64 = 1000 * format.MegaByte
 )
 
+// isHuggingFaceURL returns true if the URL is from a HuggingFace domain.
+// This includes:
+//   - huggingface.co (main domain)
+//   - *.huggingface.co (subdomains like cdn-lfs.huggingface.co)
+//   - hf.co (shortlink domain)
+//   - *.hf.co (CDN domains like cdn-lfs.hf.co, cdn-lfs3.hf.co)
+func isHuggingFaceURL(u *url.URL) bool {
+	if u == nil {
+		return false
+	}
+	host := strings.ToLower(u.Hostname())
+	return host == "huggingface.co" ||
+		strings.HasSuffix(host, ".huggingface.co") ||
+		host == "hf.co" ||
+		strings.HasSuffix(host, ".hf.co")
+}
+
+// getNumDownloadParts returns the number of concurrent download parts to use
+// for the given URL. HuggingFace URLs use reduced concurrency (default 4) to
+// avoid triggering rate limits. This can be overridden via the OLLAMA_HF_CONCURRENCY
+// environment variable. For non-HuggingFace URLs, returns the standard concurrency (16).
+func getNumDownloadParts(u *url.URL) int {
+	if isHuggingFaceURL(u) {
+		if v := os.Getenv("OLLAMA_HF_CONCURRENCY"); v != "" {
+			if n, err := strconv.Atoi(v); err == nil && n > 0 {
+				return n
+			}
+		}
+		return numHFDownloadParts
+	}
+	return numDownloadParts
+}
+
 func (p *blobDownloadPart) Name() string {
 	return strings.Join([]string{
 		p.blobDownload.Name, "partial", strconv.Itoa(p.N),
@@ -273,7 +308,11 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 	}
 
 	g, inner := errgroup.WithContext(ctx)
-	g.SetLimit(numDownloadParts)
+	concurrency := getNumDownloadParts(directURL)
+	if concurrency != numDownloadParts {
+		slog.Info(fmt.Sprintf("using reduced concurrency (%d) for HuggingFace download", concurrency))
+	}
+	g.SetLimit(concurrency)
 	for i := range b.Parts {
 		part := b.Parts[i]
 		if part.Completed.Load() == part.Size {
@@ -458,7 +497,7 @@ func (b *blobDownload) Wait(ctx context.Context, fn func(api.ProgressResponse))
 }
 
 type downloadOpts struct {
-	n       model.Name
+	mp      ModelPath
 	digest  string
 	regOpts *registryOptions
 	fn      func(api.ProgressResponse)
@@ -467,10 +506,10 @@ type downloadOpts struct {
 // downloadBlob downloads a blob from the registry and stores it in the blobs directory
 func downloadBlob(ctx context.Context, opts downloadOpts) (cacheHit bool, _ error) {
 	if opts.digest == "" {
-		return false, fmt.Errorf(("%s: %s"), opts.n.DisplayNamespaceModel(), "digest is empty")
+		return false, fmt.Errorf(("%s: %s"), opts.mp.GetNamespaceRepository(), "digest is empty")
 	}
 
-	fp, err := manifest.BlobsPath(opts.digest)
+	fp, err := GetBlobsPath(opts.digest)
 	if err != nil {
 		return false, err
 	}
@@ -494,8 +533,8 @@ func downloadBlob(ctx context.Context, opts downloadOpts) (cacheHit bool, _ erro
 	data, ok := blobDownloadManager.LoadOrStore(opts.digest, &blobDownload{Name: fp, Digest: opts.digest})
 	download := data.(*blobDownload)
 	if !ok {
-		requestURL := opts.n.BaseURL()
-		requestURL = requestURL.JoinPath("v2", opts.n.DisplayNamespaceModel(), "blobs", opts.digest)
+		requestURL := opts.mp.BaseURL()
+		requestURL = requestURL.JoinPath("v2", opts.mp.GetNamespaceRepository(), "blobs", opts.digest)
 		if err := download.Prepare(ctx, requestURL, opts.regOpts); err != nil {
 			blobDownloadManager.Delete(opts.digest)
 			return false, err

server/download_test.go

@@ -0,0 +1,194 @@
+package server
+
+import (
+	"net/url"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestIsHuggingFaceURL(t *testing.T) {
+	tests := []struct {
+		name     string
+		url      string
+		expected bool
+	}{
+		{
+			name:     "nil url",
+			url:      "",
+			expected: false,
+		},
+		{
+			name:     "huggingface.co main domain",
+			url:      "https://huggingface.co/some/model",
+			expected: true,
+		},
+		{
+			name:     "cdn-lfs.huggingface.co subdomain",
+			url:      "https://cdn-lfs.huggingface.co/repos/abc/123",
+			expected: true,
+		},
+		{
+			name:     "cdn-lfs3.hf.co CDN domain",
+			url:      "https://cdn-lfs3.hf.co/repos/abc/123",
+			expected: true,
+		},
+		{
+			name:     "hf.co shortlink domain",
+			url:      "https://hf.co/model",
+			expected: true,
+		},
+		{
+			name:     "uppercase HuggingFace domain",
+			url:      "https://HUGGINGFACE.CO/model",
+			expected: true,
+		},
+		{
+			name:     "mixed case HF domain",
+			url:      "https://Cdn-Lfs.HF.Co/repos",
+			expected: true,
+		},
+		{
+			name:     "ollama registry",
+			url:      "https://registry.ollama.ai/v2/library/llama3",
+			expected: false,
+		},
+		{
+			name:     "github.com",
+			url:      "https://github.com/ollama/ollama",
+			expected: false,
+		},
+		{
+			name:     "fake huggingface domain",
+			url:      "https://nothuggingface.co/model",
+			expected: false,
+		},
+		{
+			name:     "fake hf domain",
+			url:      "https://nothf.co/model",
+			expected: false,
+		},
+		{
+			name:     "huggingface in path not host",
+			url:      "https://example.com/huggingface.co/model",
+			expected: false,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			var u *url.URL
+			if tc.url != "" {
+				var err error
+				u, err = url.Parse(tc.url)
+				if err != nil {
+					t.Fatalf("failed to parse URL: %v", err)
+				}
+			}
+			got := isHuggingFaceURL(u)
+			assert.Equal(t, tc.expected, got)
+		})
+	}
+}
+
+func TestGetNumDownloadParts(t *testing.T) {
+	tests := []struct {
+		name        string
+		url         string
+		envValue    string
+		expected    int
+		description string
+	}{
+		{
+			name:        "nil url returns default",
+			url:         "",
+			envValue:    "",
+			expected:    numDownloadParts,
+			description: "nil URL should return standard concurrency",
+		},
+		{
+			name:        "ollama registry returns default",
+			url:         "https://registry.ollama.ai/v2/library/llama3",
+			envValue:    "",
+			expected:    numDownloadParts,
+			description: "Ollama registry should use standard concurrency",
+		},
+		{
+			name:        "huggingface returns reduced default",
+			url:         "https://huggingface.co/model/repo",
+			envValue:    "",
+			expected:    numHFDownloadParts,
+			description: "HuggingFace should use reduced concurrency",
+		},
+		{
+			name:        "hf.co CDN returns reduced default",
+			url:         "https://cdn-lfs3.hf.co/repos/abc/123",
+			envValue:    "",
+			expected:    numHFDownloadParts,
+			description: "HuggingFace CDN should use reduced concurrency",
+		},
+		{
+			name:        "huggingface with env override",
+			url:         "https://huggingface.co/model/repo",
+			envValue:    "2",
+			expected:    2,
+			description: "OLLAMA_HF_CONCURRENCY should override default",
+		},
+		{
+			name:        "huggingface with higher env override",
+			url:         "https://huggingface.co/model/repo",
+			envValue:    "8",
+			expected:    8,
+			description: "OLLAMA_HF_CONCURRENCY can be set higher than default",
+		},
+		{
+			name:        "huggingface with invalid env (non-numeric)",
+			url:         "https://huggingface.co/model/repo",
+			envValue:    "invalid",
+			expected:    numHFDownloadParts,
+			description: "Invalid OLLAMA_HF_CONCURRENCY should fall back to default",
+		},
+		{
+			name:        "huggingface with invalid env (zero)",
+			url:         "https://huggingface.co/model/repo",
+			envValue:    "0",
+			expected:    numHFDownloadParts,
+			description: "Zero OLLAMA_HF_CONCURRENCY should fall back to default",
+		},
+		{
+			name:        "huggingface with invalid env (negative)",
+			url:         "https://huggingface.co/model/repo",
+			envValue:    "-1",
+			expected:    numHFDownloadParts,
+			description: "Negative OLLAMA_HF_CONCURRENCY should fall back to default",
+		},
+		{
+			name:        "non-huggingface ignores env",
+			url:         "https://registry.ollama.ai/v2/library/llama3",
+			envValue:    "2",
+			expected:    numDownloadParts,
+			description: "OLLAMA_HF_CONCURRENCY should not affect non-HF URLs",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			// Set or clear the environment variable
+			if tc.envValue != "" {
+				t.Setenv("OLLAMA_HF_CONCURRENCY", tc.envValue)
+			}
+
+			var u *url.URL
+			if tc.url != "" {
+				var err error
+				u, err = url.Parse(tc.url)
+				if err != nil {
+					t.Fatalf("failed to parse URL: %v", err)
+				}
+			}
+
+			got := getNumDownloadParts(u)
+			assert.Equal(t, tc.expected, got, tc.description)
+		})
+	}
+}

server/images.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"context"
 	"crypto/sha256"
+	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -23,7 +24,6 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/fs/gguf"
-	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/model/parsers"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/template"
@@ -41,7 +41,6 @@ var (
 	errCapabilityVision     = errors.New("vision")
 	errCapabilityEmbedding  = errors.New("embedding")
 	errCapabilityThinking   = errors.New("thinking")
-	errCapabilityImage      = errors.New("image generation")
 	errInsecureProtocol     = errors.New("insecure protocol http")
 )
 
@@ -75,6 +74,12 @@ type Model struct {
 func (m *Model) Capabilities() []model.Capability {
 	capabilities := []model.Capability{}
 
+	// Check for image generation model via config capabilities
+	if slices.Contains(m.Config.Capabilities, "image") {
+		return []model.Capability{model.CapabilityImageGeneration}
+	}
+
+	// Check for completion capability
 	if m.ModelPath != "" {
 		f, err := gguf.Open(m.ModelPath)
 		if err == nil {
@@ -154,7 +159,6 @@ func (m *Model) CheckCapabilities(want ...model.Capability) error {
 		model.CapabilityVision:     errCapabilityVision,
 		model.CapabilityEmbedding:  errCapabilityEmbedding,
 		model.CapabilityThinking:   errCapabilityThinking,
-		model.CapabilityImage:      errCapabilityImage,
 	}
 
 	for _, cap := range want {
@@ -268,22 +272,44 @@ func (m *Model) String() string {
 	return modelfile.String()
 }
 
+func GetManifest(mp ModelPath) (*Manifest, string, error) {
+	fp, err := mp.GetManifestPath()
+	if err != nil {
+		return nil, "", err
+	}
+
+	f, err := os.Open(fp)
+	if err != nil {
+		return nil, "", err
+	}
+	defer f.Close()
+
+	sha256sum := sha256.New()
+
+	var manifest Manifest
+	if err := json.NewDecoder(io.TeeReader(f, sha256sum)).Decode(&manifest); err != nil {
+		return nil, "", err
+	}
+
+	return &manifest, hex.EncodeToString(sha256sum.Sum(nil)), nil
+}
+
 func GetModel(name string) (*Model, error) {
-	n := model.ParseName(name)
-	mf, err := manifest.ParseNamedManifest(n)
+	mp := ParseModelPath(name)
+	manifest, digest, err := GetManifest(mp)
 	if err != nil {
 		return nil, err
 	}
 
-	m := &Model{
-		Name:      n.String(),
-		ShortName: n.DisplayShortest(),
-		Digest:    mf.Digest(),
+	model := &Model{
+		Name:      mp.GetFullTagname(),
+		ShortName: mp.GetShortTagname(),
+		Digest:    digest,
 		Template:  template.DefaultTemplate,
 	}
 
-	if mf.Config.Digest != "" {
-		filename, err := manifest.BlobsPath(mf.Config.Digest)
+	if manifest.Config.Digest != "" {
+		filename, err := GetBlobsPath(manifest.Config.Digest)
 		if err != nil {
 			return nil, err
 		}
@@ -294,29 +320,29 @@ func GetModel(name string) (*Model, error) {
 		}
 		defer configFile.Close()
 
-		if err := json.NewDecoder(configFile).Decode(&m.Config); err != nil {
+		if err := json.NewDecoder(configFile).Decode(&model.Config); err != nil {
 			return nil, err
 		}
 	}
 
-	for _, layer := range mf.Layers {
-		filename, err := manifest.BlobsPath(layer.Digest)
+	for _, layer := range manifest.Layers {
+		filename, err := GetBlobsPath(layer.Digest)
 		if err != nil {
 			return nil, err
 		}
 
 		switch layer.MediaType {
 		case "application/vnd.ollama.image.model":
-			m.ModelPath = filename
-			m.ParentModel = layer.From
+			model.ModelPath = filename
+			model.ParentModel = layer.From
 		case "application/vnd.ollama.image.embed":
 			// Deprecated in versions  > 0.1.2
 			// TODO: remove this warning in a future version
 			slog.Info("WARNING: model contains embeddings, but embeddings in modelfiles have been deprecated and will be ignored.")
 		case "application/vnd.ollama.image.adapter":
-			m.AdapterPaths = append(m.AdapterPaths, filename)
+			model.AdapterPaths = append(model.AdapterPaths, filename)
 		case "application/vnd.ollama.image.projector":
-			m.ProjectorPaths = append(m.ProjectorPaths, filename)
+			model.ProjectorPaths = append(model.ProjectorPaths, filename)
 		case "application/vnd.ollama.image.prompt",
 			"application/vnd.ollama.image.template":
 			bts, err := os.ReadFile(filename)
@@ -324,7 +350,7 @@ func GetModel(name string) (*Model, error) {
 				return nil, err
 			}
 
-			m.Template, err = template.Parse(string(bts))
+			model.Template, err = template.Parse(string(bts))
 			if err != nil {
 				return nil, err
 			}
@@ -334,7 +360,7 @@ func GetModel(name string) (*Model, error) {
 				return nil, err
 			}
 
-			m.System = string(bts)
+			model.System = string(bts)
 		case "application/vnd.ollama.image.params":
 			params, err := os.Open(filename)
 			if err != nil {
@@ -343,7 +369,7 @@ func GetModel(name string) (*Model, error) {
 			defer params.Close()
 
 			// parse model options parameters into a map so that we can see which fields have been specified explicitly
-			if err = json.NewDecoder(params).Decode(&m.Options); err != nil {
+			if err = json.NewDecoder(params).Decode(&model.Options); err != nil {
 				return nil, err
 			}
 		case "application/vnd.ollama.image.messages":
@@ -353,7 +379,7 @@ func GetModel(name string) (*Model, error) {
 			}
 			defer msgs.Close()
 
-			if err = json.NewDecoder(msgs).Decode(&m.Messages); err != nil {
+			if err = json.NewDecoder(msgs).Decode(&model.Messages); err != nil {
 				return nil, err
 			}
 		case "application/vnd.ollama.image.license":
@@ -361,11 +387,11 @@ func GetModel(name string) (*Model, error) {
 			if err != nil {
 				return nil, err
 			}
-			m.License = append(m.License, string(bts))
+			model.License = append(model.License, string(bts))
 		}
 	}
 
-	return m, nil
+	return model, nil
 }
 
 func CopyModel(src, dst model.Name) error {
@@ -380,7 +406,7 @@ func CopyModel(src, dst model.Name) error {
 		return nil
 	}
 
-	manifests, err := manifest.Path()
+	manifests, err := GetManifestPath()
 	if err != nil {
 		return err
 	}
@@ -409,7 +435,7 @@ func CopyModel(src, dst model.Name) error {
 
 func deleteUnusedLayers(deleteMap map[string]struct{}) error {
 	// Ignore corrupt manifests to avoid blocking deletion of layers that are freshly orphaned
-	manifests, err := manifest.Manifests(true)
+	manifests, err := Manifests(true)
 	if err != nil {
 		return err
 	}
@@ -424,7 +450,7 @@ func deleteUnusedLayers(deleteMap map[string]struct{}) error {
 
 	// only delete the files which are still in the deleteMap
 	for k := range deleteMap {
-		fp, err := manifest.BlobsPath(k)
+		fp, err := GetBlobsPath(k)
 		if err != nil {
 			slog.Info(fmt.Sprintf("couldn't get file path for '%s': %v", k, err))
 			continue
@@ -440,7 +466,7 @@ func deleteUnusedLayers(deleteMap map[string]struct{}) error {
 
 func PruneLayers() error {
 	deleteMap := make(map[string]struct{})
-	p, err := manifest.BlobsPath("")
+	p, err := GetBlobsPath("")
 	if err != nil {
 		return err
 	}
@@ -455,9 +481,9 @@ func PruneLayers() error {
 		name := blob.Name()
 		name = strings.ReplaceAll(name, "-", ":")
 
-		_, err := manifest.BlobsPath(name)
+		_, err := GetBlobsPath(name)
 		if err != nil {
-			if errors.Is(err, manifest.ErrInvalidDigestFormat) {
+			if errors.Is(err, ErrInvalidDigestFormat) {
 				// remove invalid blobs (e.g. partial downloads)
 				if err := os.Remove(filepath.Join(p, blob.Name())); err != nil {
 					slog.Error("couldn't remove blob", "blob", blob.Name(), "error", err)
@@ -482,30 +508,63 @@ func PruneLayers() error {
 	return nil
 }
 
+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
+}
+
 func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
-	n := model.ParseName(name)
+	mp := ParseModelPath(name)
 	fn(api.ProgressResponse{Status: "retrieving manifest"})
 
-	if n.ProtocolScheme == "http" && !regOpts.Insecure {
+	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
 		return errInsecureProtocol
 	}
 
-	mf, err := manifest.ParseNamedManifest(n)
+	manifest, _, err := GetManifest(mp)
 	if err != nil {
 		fn(api.ProgressResponse{Status: "couldn't retrieve manifest"})
 		return err
 	}
 
-	var layers []manifest.Layer
-	layers = append(layers, mf.Layers...)
-	if mf.Config.Digest != "" {
-		layers = append(layers, mf.Config)
+	var layers []Layer
+	layers = append(layers, manifest.Layers...)
+	if manifest.Config.Digest != "" {
+		layers = append(layers, manifest.Config)
 	}
 
 	// Use fast transfer for models with tensor layers (many small blobs)
 	if hasTensorLayers(layers) {
 		// Read raw manifest JSON to preserve tensor metadata fields
-		manifestPath, err := manifest.PathForName(n)
+		manifestPath, err := mp.GetManifestPath()
 		if err != nil {
 			return err
 		}
@@ -513,7 +572,7 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		if err != nil {
 			return err
 		}
-		if err := pushWithTransfer(ctx, n, layers, manifestJSON, regOpts, fn); err != nil {
+		if err := pushWithTransfer(ctx, mp, layers, manifestJSON, regOpts, fn); err != nil {
 			return err
 		}
 		fn(api.ProgressResponse{Status: "success"})
@@ -521,17 +580,17 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	}
 
 	for _, layer := range layers {
-		if err := uploadBlob(ctx, n, layer, regOpts, fn); err != nil {
+		if err := uploadBlob(ctx, mp, layer, regOpts, fn); err != nil {
 			slog.Info(fmt.Sprintf("error uploading blob: %v", err))
 			return err
 		}
 	}
 
 	fn(api.ProgressResponse{Status: "pushing manifest"})
-	requestURL := n.BaseURL()
-	requestURL = requestURL.JoinPath("v2", n.DisplayNamespaceModel(), "manifests", n.Tag)
+	requestURL := mp.BaseURL()
+	requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
 
-	manifestJSON, err := json.Marshal(mf)
+	manifestJSON, err := json.Marshal(manifest)
 	if err != nil {
 		return err
 	}
@@ -550,44 +609,44 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 }
 
 func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
-	n := model.ParseName(name)
+	mp := ParseModelPath(name)
 
 	// build deleteMap to prune unused layers
 	deleteMap := make(map[string]struct{})
-	existingMf, err := manifest.ParseNamedManifest(n)
+	manifest, _, err := GetManifest(mp)
 	if errors.Is(err, os.ErrNotExist) {
 		// noop
 	} else if err != nil {
 		slog.Warn("pulling model with bad existing manifest", "name", name, "error", err)
 	} else {
-		for _, l := range existingMf.Layers {
+		for _, l := range manifest.Layers {
 			deleteMap[l.Digest] = struct{}{}
 		}
-		if existingMf.Config.Digest != "" {
-			deleteMap[existingMf.Config.Digest] = struct{}{}
+		if manifest.Config.Digest != "" {
+			deleteMap[manifest.Config.Digest] = struct{}{}
 		}
 	}
 
-	if n.ProtocolScheme == "http" && !regOpts.Insecure {
+	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
 		return errInsecureProtocol
 	}
 
 	fn(api.ProgressResponse{Status: "pulling manifest"})
 
-	mf, err := pullModelManifest(ctx, n, regOpts)
+	manifest, err = pullModelManifest(ctx, mp, regOpts)
 	if err != nil {
 		return fmt.Errorf("pull model manifest: %s", err)
 	}
 
-	var layers []manifest.Layer
-	layers = append(layers, mf.Layers...)
-	if mf.Config.Digest != "" {
-		layers = append(layers, mf.Config)
+	var layers []Layer
+	layers = append(layers, manifest.Layers...)
+	if manifest.Config.Digest != "" {
+		layers = append(layers, manifest.Config)
 	}
 
 	// Use fast transfer for models with tensor layers (many small blobs)
 	if hasTensorLayers(layers) {
-		if err := pullWithTransfer(ctx, n, layers, mf, regOpts, fn); err != nil {
+		if err := pullWithTransfer(ctx, mp, layers, manifest, regOpts, fn); err != nil {
 			return err
 		}
 		fn(api.ProgressResponse{Status: "success"})
@@ -597,7 +656,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	skipVerify := make(map[string]bool)
 	for _, layer := range layers {
 		cacheHit, err := downloadBlob(ctx, downloadOpts{
-			n:       n,
+			mp:      mp,
 			digest:  layer.Digest,
 			regOpts: regOpts,
 			fn:      fn,
@@ -616,7 +675,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		}
 		if err := verifyBlob(layer.Digest); err != nil {
 			if errors.Is(err, errDigestMismatch) {
-				fp, err := manifest.BlobsPath(layer.Digest)
+				fp, err := GetBlobsPath(layer.Digest)
 				if err != nil {
 					return err
 				}
@@ -631,16 +690,16 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	for _, layer := range layers {
 		delete(deleteMap, layer.Digest)
 	}
-	delete(deleteMap, mf.Config.Digest)
+	delete(deleteMap, manifest.Config.Digest)
 
 	fn(api.ProgressResponse{Status: "writing manifest"})
 
-	manifestJSON, err := json.Marshal(mf)
+	manifestJSON, err := json.Marshal(manifest)
 	if err != nil {
 		return err
 	}
 
-	fp, err := manifest.PathForName(n)
+	fp, err := mp.GetManifestPath()
 	if err != nil {
 		return err
 	}
@@ -667,9 +726,9 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 }
 
 // hasTensorLayers checks if any layer has tensor media type.
-func hasTensorLayers(layers []manifest.Layer) bool {
+func hasTensorLayers(layers []Layer) bool {
 	for _, layer := range layers {
-		if layer.MediaType == manifest.MediaTypeImageTensor {
+		if layer.MediaType == MediaTypeImageTensor {
 			return true
 		}
 	}
@@ -677,7 +736,7 @@ func hasTensorLayers(layers []manifest.Layer) bool {
 }
 
 // pullWithTransfer uses the simplified x/transfer package for downloading blobs.
-func pullWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer, mf *manifest.Manifest, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
+func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifest *Manifest, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
 	blobs := make([]transfer.Blob, len(layers))
 	for i, layer := range layers {
 		blobs[i] = transfer.Blob{
@@ -686,12 +745,12 @@ func pullWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer
 		}
 	}
 
-	destDir, err := manifest.BlobsPath("")
+	destDir, err := GetBlobsPath("")
 	if err != nil {
 		return err
 	}
 
-	base := n.BaseURL()
+	base := mp.BaseURL()
 	if base.Scheme != "http" && regOpts != nil && regOpts.Insecure {
 		base.Scheme = "http"
 	}
@@ -716,14 +775,14 @@ func pullWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer
 			Realm:   challenge.Realm,
 			Service: challenge.Service,
 			Scope:   challenge.Scope,
-		}, base.Host)
+		})
 	}
 
 	if err := transfer.Download(ctx, transfer.DownloadOptions{
 		Blobs:      blobs,
 		BaseURL:    baseURL,
 		DestDir:    destDir,
-		Repository: n.DisplayNamespaceModel(),
+		Repository: mp.GetNamespaceRepository(),
 		Progress:   progress,
 		Token:      regOpts.Token,
 		GetToken:   getToken,
@@ -734,12 +793,12 @@ func pullWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer
 
 	// Write manifest
 	fn(api.ProgressResponse{Status: "writing manifest"})
-	manifestJSON, err := json.Marshal(mf)
+	manifestJSON, err := json.Marshal(manifest)
 	if err != nil {
 		return err
 	}
 
-	fp, err := manifest.PathForName(n)
+	fp, err := mp.GetManifestPath()
 	if err != nil {
 		return err
 	}
@@ -751,7 +810,7 @@ func pullWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer
 }
 
 // pushWithTransfer uses the simplified x/transfer package for uploading blobs and manifest.
-func pushWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer, manifestJSON []byte, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
+func pushWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifestJSON []byte, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
 	blobs := make([]transfer.Blob, len(layers))
 	for i, layer := range layers {
 		blobs[i] = transfer.Blob{
@@ -761,12 +820,12 @@ func pushWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer
 		}
 	}
 
-	srcDir, err := manifest.BlobsPath("")
+	srcDir, err := GetBlobsPath("")
 	if err != nil {
 		return err
 	}
 
-	base := n.BaseURL()
+	base := mp.BaseURL()
 	if base.Scheme != "http" && regOpts != nil && regOpts.Insecure {
 		base.Scheme = "http"
 	}
@@ -791,7 +850,7 @@ func pushWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer
 			Realm:   challenge.Realm,
 			Service: challenge.Service,
 			Scope:   challenge.Scope,
-		}, base.Host)
+		})
 	}
 
 	return transfer.Upload(ctx, transfer.UploadOptions{
@@ -803,13 +862,13 @@ func pushWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer
 		GetToken:    getToken,
 		Logger:      slog.Default(),
 		Manifest:    manifestJSON,
-		ManifestRef: n.Tag,
-		Repository:  n.DisplayNamespaceModel(),
+		ManifestRef: mp.Tag,
+		Repository:  mp.GetNamespaceRepository(),
 	})
 }
 
-func pullModelManifest(ctx context.Context, n model.Name, regOpts *registryOptions) (*manifest.Manifest, error) {
-	requestURL := n.BaseURL().JoinPath("v2", n.DisplayNamespaceModel(), "manifests", n.Tag)
+func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*Manifest, error) {
+	requestURL := mp.BaseURL().JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
 
 	headers := make(http.Header)
 	headers.Set("Accept", "application/vnd.docker.distribution.manifest.v2+json")
@@ -819,7 +878,7 @@ func pullModelManifest(ctx context.Context, n model.Name, regOpts *registryOptio
 	}
 	defer resp.Body.Close()
 
-	var m manifest.Manifest
+	var m Manifest
 	if err := json.NewDecoder(resp.Body).Decode(&m); err != nil {
 		return nil, err
 	}
@@ -857,7 +916,7 @@ func makeRequestWithRetry(ctx context.Context, method string, requestURL *url.UR
 
 			// Handle authentication error with one retry
 			challenge := parseRegistryChallenge(resp.Header.Get("www-authenticate"))
-			token, err := getAuthorizationToken(ctx, challenge, requestURL.Host)
+			token, err := getAuthorizationToken(ctx, challenge)
 			if err != nil {
 				return nil, err
 			}
@@ -981,7 +1040,7 @@ func parseRegistryChallenge(authStr string) registryChallenge {
 var errDigestMismatch = errors.New("digest mismatch, file must be downloaded again")
 
 func verifyBlob(digest string) error {
-	fp, err := manifest.BlobsPath(digest)
+	fp, err := GetBlobsPath(digest)
 	if err != nil {
 		return err
 	}

server/images_test.go

@@ -54,16 +54,7 @@ func TestModelCapabilities(t *testing.T) {
 					Capabilities: []string{"image"},
 				},
 			},
-			expectedCaps: []model.Capability{model.CapabilityImage},
-		},
-		{
-			name: "model with image and vision capability (image editing)",
-			model: Model{
-				Config: model.ConfigV2{
-					Capabilities: []string{"image", "vision"},
-				},
-			},
-			expectedCaps: []model.Capability{model.CapabilityImage, model.CapabilityVision},
+			expectedCaps: []model.Capability{model.CapabilityImageGeneration},
 		},
 		{
 			name: "model with completion capability",
@@ -251,24 +242,6 @@ func TestModelCheckCapabilities(t *testing.T) {
 			checkCaps:      []model.Capability{"unknown"},
 			expectedErrMsg: "unknown capability",
 		},
-		{
-			name: "model missing image generation capability",
-			model: Model{
-				ModelPath: completionModelPath,
-				Template:  chatTemplate,
-			},
-			checkCaps:      []model.Capability{model.CapabilityImage},
-			expectedErrMsg: "does not support image generation",
-		},
-		{
-			name: "model with image generation capability",
-			model: Model{
-				Config: model.ConfigV2{
-					Capabilities: []string{"image"},
-				},
-			},
-			checkCaps: []model.Capability{model.CapabilityImage},
-		},
 	}
 
 	for _, tt := range tests {

server/layer.go

@@ -1,4 +1,4 @@
-package manifest
+package server
 
 import (
 	"crypto/sha256"
@@ -14,7 +14,7 @@ type Layer struct {
 	Size      int64  `json:"size"`
 	From      string `json:"from,omitempty"`
 	Name      string `json:"name,omitempty"` // tensor name, e.g., "text_encoder/model.embed_tokens.weight"
-	Status    string `json:"-"`
+	status    string
 }
 
 const (
@@ -22,7 +22,7 @@ const (
 )
 
 func NewLayer(r io.Reader, mediatype string) (Layer, error) {
-	blobs, err := BlobsPath("")
+	blobs, err := GetBlobsPath("")
 	if err != nil {
 		return Layer{}, err
 	}
@@ -45,7 +45,7 @@ func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 	}
 
 	digest := fmt.Sprintf("sha256:%x", sha256sum.Sum(nil))
-	blob, err := BlobsPath(digest)
+	blob, err := GetBlobsPath(digest)
 	if err != nil {
 		return Layer{}, err
 	}
@@ -65,7 +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
 }
 
@@ -74,7 +74,7 @@ func NewLayerFromLayer(digest, mediatype, from string) (Layer, error) {
 		return Layer{}, errors.New("creating new layer from layer with empty digest")
 	}
 
-	blob, err := BlobsPath(digest)
+	blob, err := GetBlobsPath(digest)
 	if err != nil {
 		return Layer{}, err
 	}
@@ -89,7 +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
 }
 
@@ -98,7 +98,7 @@ func (l *Layer) Open() (io.ReadSeekCloser, error) {
 		return nil, errors.New("opening layer with empty digest")
 	}
 
-	blob, err := BlobsPath(l.Digest)
+	blob, err := GetBlobsPath(l.Digest)
 	if err != nil {
 		return nil, err
 	}
@@ -126,7 +126,7 @@ func (l *Layer) Remove() error {
 		}
 	}
 
-	blob, err := BlobsPath(l.Digest)
+	blob, err := GetBlobsPath(l.Digest)
 	if err != nil {
 		return err
 	}

server/manifest.go

@@ -1,9 +1,10 @@
-package manifest
+package server
 
 import (
 	"crypto/sha256"
 	"encoding/hex"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -32,38 +33,12 @@ 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 := Path()
+	manifests, err := GetManifestPath()
 	if err != nil {
 		return err
 	}
@@ -95,11 +70,11 @@ func (m *Manifest) RemoveLayers() error {
 		if _, used := inUse[layer.Digest]; used {
 			continue
 		}
-		blob, err := BlobsPath(layer.Digest)
+		blob, err := GetBlobsPath(layer.Digest)
 		if err != nil {
 			return err
 		}
-		if err := os.Remove(blob); os.IsNotExist(err) {
+		if err := os.Remove(blob); errors.Is(err, os.ErrNotExist) {
 			slog.Debug("layer does not exist", "digest", layer.Digest)
 		} else if err != nil {
 			return err
@@ -114,7 +89,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 		return nil, model.Unqualified(n)
 	}
 
-	manifests, err := Path()
+	manifests, err := GetManifestPath()
 	if err != nil {
 		return nil, err
 	}
@@ -146,7 +121,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 }
 
 func WriteManifest(name model.Name, config Layer, layers []Layer) error {
-	manifests, err := Path()
+	manifests, err := GetManifestPath()
 	if err != nil {
 		return err
 	}
@@ -173,7 +148,7 @@ func WriteManifest(name model.Name, config Layer, layers []Layer) error {
 }
 
 func Manifests(continueOnError bool) (map[model.Name]*Manifest, error) {
-	manifests, err := Path()
+	manifests, err := GetManifestPath()
 	if err != nil {
 		return nil, err
 	}

server/manifest_test.go

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