pr feedback

This change adds a new MLX based runner which includes:

  * Method-based MLX bindings
  * Subprocess-based MLX runner (x/mlxrunner)
  * KV cache with tree management
  * A basic sampler

The GLM4-MoE-Lite model has been ported to use the new bindings.

---------

Co-authored-by: Michael Yang <git@mxy.ng>

.github/workflows/release.yaml

@@ -337,6 +337,7 @@ jobs:
           name: bundles-windows
           path: |
             dist/*.zip
+            dist/*.ps1
             dist/OllamaSetup.exe
 
   linux-build:
@@ -514,6 +515,9 @@ jobs:
       - name: Log dist contents
         run: |
           ls -l dist/
+      - name: Copy install scripts to dist
+        run: |
+          cp scripts/install.sh dist/install.sh
       - name: Generate checksum file
         run: find . -type f -not -name 'sha256sum.txt' | xargs sha256sum | tee sha256sum.txt
         working-directory: dist
@@ -536,7 +540,7 @@ jobs:
       - name: Upload release artifacts
         run: |
           pids=()
-          for payload in dist/*.txt dist/*.zip dist/*.tgz dist/*.tar.zst dist/*.exe dist/*.dmg ; do
+          for payload in dist/*.txt dist/*.zip dist/*.tgz dist/*.tar.zst dist/*.exe dist/*.dmg dist/*.ps1 dist/*.sh ; do
             echo "Uploading $payload"
             gh release upload ${GITHUB_REF_NAME} $payload --clobber &
             pids[$!]=$!

.github/workflows/test-install.yaml

@@ -0,0 +1,22 @@
+name: test-install
+
+on:
+  pull_request:
+    paths:
+      - 'scripts/install.sh'
+      - '.github/workflows/test-install.yaml'
+
+jobs:
+  test:
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Run install script
+        run: sh ./scripts/install.sh
+        env:
+          OLLAMA_NO_START: 1 # do not start app
+      - name: Verify ollama is available
+        run: ollama --version

CMakeLists.txt

@@ -182,7 +182,7 @@ option(MLX_ENGINE "Enable MLX backend" OFF)
 
 if(MLX_ENGINE)
     message(STATUS "Setting up MLX (this takes a while...)")
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/ml/backend/mlx)
+    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/imagegen/mlx)
 
     # Find CUDA toolkit if MLX is built with CUDA support
     find_package(CUDAToolkit)
@@ -216,4 +216,4 @@ if(MLX_ENGINE)
                 COMPONENT MLX)
         endif()
     endif()
-endif()
+endif()

Dockerfile

@@ -147,7 +147,7 @@ ARG PARALLEL
 WORKDIR /go/src/github.com/ollama/ollama
 COPY CMakeLists.txt CMakePresets.json .
 COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-COPY x/ml/backend/mlx x/ml/backend/mlx
+COPY x/imagegen/mlx x/imagegen/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 $(uname -m) in x86_64) echo amd64 ;; aarch64) echo arm64 ;; esac).tar.gz | tar xz -C /usr/local
@@ -169,8 +169,10 @@ COPY . .
 RUN git clone --depth 1 --branch "$(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
+ENV CGO_CFLAGS="${CGO_CFLAGS} -I/go/src/github.com/ollama/ollama/build/_deps/mlx-c-src"
+ENV CGO_CXXFLAGS="${CGO_CXXFLAGS}"
 RUN --mount=type=cache,target=/root/.cache/go-build \
     go build -tags mlx -trimpath -buildmode=pie -o /bin/ollama .
 

README.md

@@ -358,6 +358,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Odin Runes](https://github.com/leonid20000/OdinRunes)
 - [LLM-X](https://github.com/mrdjohnson/llm-x) (Progressive Web App)
 - [AnythingLLM (Docker + MacOs/Windows/Linux native app)](https://github.com/Mintplex-Labs/anything-llm)
+- [Screenpipe](https://github.com/mediar-ai/screenpipe) (24/7 screen & mic recording with AI-powered search, uses Ollama for local LLM features)
 - [Ollama Basic Chat: Uses HyperDiv Reactive UI](https://github.com/rapidarchitect/ollama_basic_chat)
 - [Ollama-chats RPG](https://github.com/drazdra/ollama-chats)
 - [IntelliBar](https://intellibar.app/) (AI-powered assistant for macOS)
@@ -465,6 +466,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Clueless](https://github.com/KashyapTan/clueless) (Open Source & Local Cluely: A desktop application LLM assistant to help you talk to anything on your screen using locally served Ollama models. Also undetectable to screenshare)
 - [ollama-co2](https://github.com/carbonatedWaterOrg/ollama-co2) (FastAPI web interface for monitoring and managing local and remote Ollama servers with real-time model monitoring and concurrent downloads)
 - [Hillnote](https://hillnote.com) (A Markdown-first workspace designed to supercharge your AI workflow. Create documents ready to integrate with Claude, ChatGPT, Gemini, Cursor, and more - all while keeping your work on your device.)
+- [Stakpak](https://github.com/stakpak/agent) (An open source, vendor neutral DevOps agent that works with any model, and any stack, for teams who just want to ship)
 
 ### Cloud
 
@@ -558,7 +560,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaFarm for Go](https://github.com/presbrey/ollamafarm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
-- [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
+- [Ollama for Ruby](https://github.com/crmne/ruby_llm)
 - [Ollama-rs for Rust](https://github.com/pepperoni21/ollama-rs)
 - [Ollama-hpp for C++](https://github.com/jmont-dev/ollama-hpp)
 - [Ollama4j for Java](https://github.com/ollama4j/ollama4j)

anthropic/anthropic.go

@@ -211,6 +211,7 @@ type MessageDelta struct {
 
 // DeltaUsage contains cumulative token usage
 type DeltaUsage struct {
+	InputTokens  int `json:"input_tokens"`
 	OutputTokens int `json:"output_tokens"`
 }
 
@@ -517,24 +518,26 @@ func mapStopReason(reason string, hasToolCalls bool) string {
 
 // StreamConverter manages state for converting Ollama streaming responses to Anthropic format
 type StreamConverter struct {
-	ID              string
-	Model           string
-	firstWrite      bool
-	contentIndex    int
-	inputTokens     int
-	outputTokens    int
-	thinkingStarted bool
-	thinkingDone    bool
-	textStarted     bool
-	toolCallsSent   map[string]bool
+	ID                   string
+	Model                string
+	firstWrite           bool
+	contentIndex         int
+	inputTokens          int
+	outputTokens         int
+	estimatedInputTokens int // Estimated tokens from request (used when actual metrics are 0)
+	thinkingStarted      bool
+	thinkingDone         bool
+	textStarted          bool
+	toolCallsSent        map[string]bool
 }
 
-func NewStreamConverter(id, model string) *StreamConverter {
+func NewStreamConverter(id, model string, estimatedInputTokens int) *StreamConverter {
 	return &StreamConverter{
-		ID:            id,
-		Model:         model,
-		firstWrite:    true,
-		toolCallsSent: make(map[string]bool),
+		ID:                   id,
+		Model:                model,
+		firstWrite:           true,
+		estimatedInputTokens: estimatedInputTokens,
+		toolCallsSent:        make(map[string]bool),
 	}
 }
 
@@ -550,7 +553,11 @@ func (c *StreamConverter) Process(r api.ChatResponse) []StreamEvent {
 
 	if c.firstWrite {
 		c.firstWrite = false
+		// Use actual metrics if available, otherwise use estimate
 		c.inputTokens = r.Metrics.PromptEvalCount
+		if c.inputTokens == 0 && c.estimatedInputTokens > 0 {
+			c.inputTokens = c.estimatedInputTokens
+		}
 
 		events = append(events, StreamEvent{
 			Event: "message_start",
@@ -721,6 +728,7 @@ func (c *StreamConverter) Process(r api.ChatResponse) []StreamEvent {
 			})
 		}
 
+		c.inputTokens = r.Metrics.PromptEvalCount
 		c.outputTokens = r.Metrics.EvalCount
 		stopReason := mapStopReason(r.DoneReason, len(c.toolCallsSent) > 0)
 
@@ -732,6 +740,7 @@ func (c *StreamConverter) Process(r api.ChatResponse) []StreamEvent {
 					StopReason: stopReason,
 				},
 				Usage: DeltaUsage{
+					InputTokens:  c.inputTokens,
 					OutputTokens: c.outputTokens,
 				},
 			},
@@ -776,3 +785,117 @@ func mapToArgs(m map[string]any) api.ToolCallFunctionArguments {
 	}
 	return args
 }
+
+// CountTokensRequest represents an Anthropic count_tokens request
+type CountTokensRequest struct {
+	Model    string          `json:"model"`
+	Messages []MessageParam  `json:"messages"`
+	System   any             `json:"system,omitempty"`
+	Tools    []Tool          `json:"tools,omitempty"`
+	Thinking *ThinkingConfig `json:"thinking,omitempty"`
+}
+
+// EstimateInputTokens estimates input tokens from a MessagesRequest (reuses CountTokensRequest logic)
+func EstimateInputTokens(req MessagesRequest) int {
+	return estimateTokens(CountTokensRequest{
+		Model:    req.Model,
+		Messages: req.Messages,
+		System:   req.System,
+		Tools:    req.Tools,
+		Thinking: req.Thinking,
+	})
+}
+
+// CountTokensResponse represents an Anthropic count_tokens response
+type CountTokensResponse struct {
+	InputTokens int `json:"input_tokens"`
+}
+
+// estimateTokens returns a rough estimate of tokens (len/4).
+// TODO: Replace with actual tokenization via Tokenize API for accuracy.
+// Current len/4 heuristic is a rough approximation (~4 chars/token average).
+func estimateTokens(req CountTokensRequest) int {
+	var totalLen int
+
+	// Count system prompt
+	if req.System != nil {
+		totalLen += countAnyContent(req.System)
+	}
+
+	// Count messages
+	for _, msg := range req.Messages {
+		// Count role (always present)
+		totalLen += len(msg.Role)
+		// Count content
+		contentLen := countAnyContent(msg.Content)
+		totalLen += contentLen
+	}
+
+	for _, tool := range req.Tools {
+		totalLen += len(tool.Name) + len(tool.Description) + len(tool.InputSchema)
+	}
+
+	// Return len/4 as rough token estimate, minimum 1 if there's any content
+	tokens := totalLen / 4
+	if tokens == 0 && (len(req.Messages) > 0 || req.System != nil) {
+		tokens = 1
+	}
+	return tokens
+}
+
+func countAnyContent(content any) int {
+	if content == nil {
+		return 0
+	}
+
+	switch c := content.(type) {
+	case string:
+		return len(c)
+	case []any:
+		total := 0
+		for _, block := range c {
+			total += countContentBlock(block)
+		}
+		return total
+	default:
+		if data, err := json.Marshal(content); err == nil {
+			return len(data)
+		}
+		return 0
+	}
+}
+
+func countContentBlock(block any) int {
+	blockMap, ok := block.(map[string]any)
+	if !ok {
+		if s, ok := block.(string); ok {
+			return len(s)
+		}
+		return 0
+	}
+
+	total := 0
+	blockType, _ := blockMap["type"].(string)
+
+	if text, ok := blockMap["text"].(string); ok {
+		total += len(text)
+	}
+
+	if thinking, ok := blockMap["thinking"].(string); ok {
+		total += len(thinking)
+	}
+
+	if blockType == "tool_use" {
+		if data, err := json.Marshal(blockMap); err == nil {
+			total += len(data)
+		}
+	}
+
+	if blockType == "tool_result" {
+		if data, err := json.Marshal(blockMap); err == nil {
+			total += len(data)
+		}
+	}
+
+	return total
+}

anthropic/anthropic_test.go

@@ -321,8 +321,6 @@ func TestFromMessagesRequest_WithThinking(t *testing.T) {
 	}
 }
 
-// TestFromMessagesRequest_ThinkingOnlyBlock verifies that messages containing only
-// a thinking block (no text, images, or tool calls) are preserved and not dropped.
 func TestFromMessagesRequest_ThinkingOnlyBlock(t *testing.T) {
 	req := MessagesRequest{
 		Model:     "test-model",
@@ -605,7 +603,7 @@ func TestGenerateMessageID(t *testing.T) {
 }
 
 func TestStreamConverter_Basic(t *testing.T) {
-	conv := NewStreamConverter("msg_123", "test-model")
+	conv := NewStreamConverter("msg_123", "test-model", 0)
 
 	// First chunk
 	resp1 := api.ChatResponse{
@@ -642,7 +640,7 @@ func TestStreamConverter_Basic(t *testing.T) {
 		},
 		Done:       true,
 		DoneReason: "stop",
-		Metrics:    api.Metrics{EvalCount: 5},
+		Metrics:    api.Metrics{PromptEvalCount: 10, EvalCount: 5},
 	}
 
 	events2 := conv.Process(resp2)
@@ -650,6 +648,24 @@ func TestStreamConverter_Basic(t *testing.T) {
 	// Should have content_block_delta, content_block_stop, message_delta, message_stop
 	hasStop := false
 	for _, e := range events2 {
+		if e.Event == "message_delta" {
+			if data, ok := e.Data.(MessageDeltaEvent); ok {
+				if data.Type != "message_delta" {
+					t.Errorf("unexpected data type: %+v", data)
+				}
+
+				if data.Delta.StopReason != "end_turn" {
+					t.Errorf("unexpected stop reason: %+v", data.Delta.StopReason)
+				}
+
+				if data.Usage.InputTokens != 10 || data.Usage.OutputTokens != 5 {
+					t.Errorf("unexpected usage: %+v", data.Usage)
+				}
+			} else {
+				t.Errorf("unexpected data: %+v", e.Data)
+			}
+		}
+
 		if e.Event == "message_stop" {
 			hasStop = true
 		}
@@ -660,7 +676,7 @@ func TestStreamConverter_Basic(t *testing.T) {
 }
 
 func TestStreamConverter_WithToolCalls(t *testing.T) {
-	conv := NewStreamConverter("msg_123", "test-model")
+	conv := NewStreamConverter("msg_123", "test-model", 0)
 
 	resp := api.ChatResponse{
 		Model: "test-model",
@@ -713,7 +729,7 @@ func TestStreamConverter_WithToolCalls(t *testing.T) {
 func TestStreamConverter_ToolCallWithUnmarshalableArgs(t *testing.T) {
 	// Test that unmarshalable arguments (like channels) are handled gracefully
 	// and don't cause a panic or corrupt stream
-	conv := NewStreamConverter("msg_123", "test-model")
+	conv := NewStreamConverter("msg_123", "test-model", 0)
 
 	// Create a channel which cannot be JSON marshaled
 	unmarshalable := make(chan int)
@@ -760,7 +776,7 @@ func TestStreamConverter_ToolCallWithUnmarshalableArgs(t *testing.T) {
 
 func TestStreamConverter_MultipleToolCallsWithMixedValidity(t *testing.T) {
 	// Test that valid tool calls still work when mixed with invalid ones
-	conv := NewStreamConverter("msg_123", "test-model")
+	conv := NewStreamConverter("msg_123", "test-model", 0)
 
 	unmarshalable := make(chan int)
 	badArgs := api.NewToolCallFunctionArguments()
@@ -824,10 +840,6 @@ func TestStreamConverter_MultipleToolCallsWithMixedValidity(t *testing.T) {
 	}
 }
 
-// TestContentBlockJSON_EmptyFieldsPresent verifies that empty text and thinking fields
-// are serialized in JSON output. The Anthropic SDK requires these fields to be present
-// (even when empty) in content_block_start events to properly accumulate streaming deltas.
-// Without these fields, the SDK throws: "TypeError: unsupported operand type(s) for +=: 'NoneType' and 'str'"
 func TestContentBlockJSON_EmptyFieldsPresent(t *testing.T) {
 	tests := []struct {
 		name     string
@@ -881,11 +893,9 @@ func TestContentBlockJSON_EmptyFieldsPresent(t *testing.T) {
 	}
 }
 
-// TestStreamConverter_ContentBlockStartIncludesEmptyFields verifies that content_block_start
-// events include the required empty fields for SDK compatibility.
 func TestStreamConverter_ContentBlockStartIncludesEmptyFields(t *testing.T) {
 	t.Run("text block start includes empty text", func(t *testing.T) {
-		conv := NewStreamConverter("msg_123", "test-model")
+		conv := NewStreamConverter("msg_123", "test-model", 0)
 
 		resp := api.ChatResponse{
 			Model:   "test-model",
@@ -919,7 +929,7 @@ func TestStreamConverter_ContentBlockStartIncludesEmptyFields(t *testing.T) {
 	})
 
 	t.Run("thinking block start includes empty thinking", func(t *testing.T) {
-		conv := NewStreamConverter("msg_123", "test-model")
+		conv := NewStreamConverter("msg_123", "test-model", 0)
 
 		resp := api.ChatResponse{
 			Model:   "test-model",
@@ -951,3 +961,105 @@ func TestStreamConverter_ContentBlockStartIncludesEmptyFields(t *testing.T) {
 		}
 	})
 }
+
+func TestEstimateTokens_SimpleMessage(t *testing.T) {
+	req := CountTokensRequest{
+		Model: "test-model",
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello, world!"},
+		},
+	}
+
+	tokens := estimateTokens(req)
+
+	// "user" (4) + "Hello, world!" (13) = 17 chars / 4 = 4 tokens
+	if tokens < 1 {
+		t.Errorf("expected at least 1 token, got %d", tokens)
+	}
+	// Sanity check: shouldn't be wildly off
+	if tokens > 10 {
+		t.Errorf("expected fewer than 10 tokens for short message, got %d", tokens)
+	}
+}
+
+func TestEstimateTokens_WithSystemPrompt(t *testing.T) {
+	req := CountTokensRequest{
+		Model:  "test-model",
+		System: "You are a helpful assistant.",
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	tokens := estimateTokens(req)
+
+	// System prompt adds to count
+	if tokens < 5 {
+		t.Errorf("expected at least 5 tokens with system prompt, got %d", tokens)
+	}
+}
+
+func TestEstimateTokens_WithTools(t *testing.T) {
+	req := CountTokensRequest{
+		Model: "test-model",
+		Messages: []MessageParam{
+			{Role: "user", Content: "What's the weather?"},
+		},
+		Tools: []Tool{
+			{
+				Name:        "get_weather",
+				Description: "Get the current weather for a location",
+				InputSchema: json.RawMessage(`{"type":"object","properties":{"location":{"type":"string"}}}`),
+			},
+		},
+	}
+
+	tokens := estimateTokens(req)
+
+	// Tools add significant content
+	if tokens < 10 {
+		t.Errorf("expected at least 10 tokens with tools, got %d", tokens)
+	}
+}
+
+func TestEstimateTokens_WithThinking(t *testing.T) {
+	req := CountTokensRequest{
+		Model: "test-model",
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type":     "thinking",
+						"thinking": "Let me think about this carefully...",
+					},
+					map[string]any{
+						"type": "text",
+						"text": "Here is my response.",
+					},
+				},
+			},
+		},
+	}
+
+	tokens := estimateTokens(req)
+
+	// Thinking content should be counted
+	if tokens < 10 {
+		t.Errorf("expected at least 10 tokens with thinking content, got %d", tokens)
+	}
+}
+
+func TestEstimateTokens_EmptyContent(t *testing.T) {
+	req := CountTokensRequest{
+		Model:    "test-model",
+		Messages: []MessageParam{},
+	}
+
+	tokens := estimateTokens(req)
+
+	if tokens != 0 {
+		t.Errorf("expected 0 tokens for empty content, got %d", tokens)
+	}
+}

api/client.go

@@ -466,3 +466,25 @@ func (c *Client) Whoami(ctx context.Context) (*UserResponse, error) {
 	}
 	return &resp, nil
 }
+
+// AliasRequest is the request body for creating or updating a model alias.
+type AliasRequest struct {
+	Alias          string `json:"alias"`
+	Target         string `json:"target"`
+	PrefixMatching bool   `json:"prefix_matching,omitempty"`
+}
+
+// SetAliasExperimental creates or updates a model alias via the experimental aliases API.
+func (c *Client) SetAliasExperimental(ctx context.Context, req *AliasRequest) error {
+	return c.do(ctx, http.MethodPost, "/api/experimental/aliases", req, nil)
+}
+
+// AliasDeleteRequest is the request body for deleting a model alias.
+type AliasDeleteRequest struct {
+	Alias string `json:"alias"`
+}
+
+// DeleteAliasExperimental deletes a model alias via the experimental aliases API.
+func (c *Client) DeleteAliasExperimental(ctx context.Context, req *AliasDeleteRequest) error {
+	return c.do(ctx, http.MethodDelete, "/api/experimental/aliases", req, nil)
+}

api/types.go

@@ -749,7 +749,7 @@ type ShowResponse struct {
 	Messages      []Message          `json:"messages,omitempty"`
 	RemoteModel   string             `json:"remote_model,omitempty"`
 	RemoteHost    string             `json:"remote_host,omitempty"`
-	ModelInfo     map[string]any     `json:"model_info,omitempty"`
+	ModelInfo     map[string]any     `json:"model_info"`
 	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
 	Tensors       []Tensor           `json:"tensors,omitempty"`
 	Capabilities  []model.Capability `json:"capabilities,omitempty"`

app/README.md

@@ -75,9 +75,9 @@ The `-dev` flag enables:
 CI builds with Xcode 14.1 for OS compatibility prior to v13.  If you want to manually build v11+ support, you can download the older Xcode [here](https://developer.apple.com/services-account/download?path=/Developer_Tools/Xcode_14.1/Xcode_14.1.xip), extract, then `mv ./Xcode.app /Applications/Xcode_14.1.0.app` then activate with:
 
 ```
-export CGO_CFLAGS=-mmacosx-version-min=12.0
-export CGO_CXXFLAGS=-mmacosx-version-min=12.0
-export CGO_LDFLAGS=-mmacosx-version-min=12.0
+export CGO_CFLAGS="-O3 -mmacosx-version-min=12.0"
+export CGO_CXXFLAGS="-O3 -mmacosx-version-min=12.0"
+export CGO_LDFLAGS="-mmacosx-version-min=12.0"
 export SDKROOT=/Applications/Xcode_14.1.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk
 export DEVELOPER_DIR=/Applications/Xcode_14.1.0.app/Contents/Developer
 ```

cmd/background_unix.go

@@ -0,0 +1,13 @@
+//go:build !windows
+
+package cmd
+
+import "syscall"
+
+// backgroundServerSysProcAttr returns SysProcAttr for running the server in the background on Unix.
+// Setpgid prevents the server from being killed when the parent process exits.
+func backgroundServerSysProcAttr() *syscall.SysProcAttr {
+	return &syscall.SysProcAttr{
+		Setpgid: true,
+	}
+}

cmd/background_windows.go

@@ -0,0 +1,12 @@
+package cmd
+
+import "syscall"
+
+// backgroundServerSysProcAttr returns SysProcAttr for running the server in the background on Windows.
+// CREATE_NO_WINDOW (0x08000000) prevents a console window from appearing.
+func backgroundServerSysProcAttr() *syscall.SysProcAttr {
+	return &syscall.SysProcAttr{
+		CreationFlags: 0x08000000,
+		HideWindow:    true,
+	}
+}

cmd/cmd.go

@@ -15,6 +15,7 @@ import (
 	"net"
 	"net/http"
 	"os"
+	"os/exec"
 	"os/signal"
 	"path/filepath"
 	"runtime"
@@ -29,12 +30,15 @@ import (
 	"github.com/containerd/console"
 	"github.com/mattn/go-runewidth"
 	"github.com/olekukonko/tablewriter"
+	"github.com/pkg/browser"
 	"github.com/spf13/cobra"
 	"golang.org/x/crypto/ssh"
 	"golang.org/x/sync/errgroup"
 	"golang.org/x/term"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/cmd/config"
+	"github.com/ollama/ollama/cmd/tui"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/parser"
@@ -51,7 +55,50 @@ import (
 	"github.com/ollama/ollama/x/imagegen"
 )
 
-const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
+func init() {
+	// Override default selectors to use Bubbletea TUI instead of raw terminal I/O.
+	config.DefaultSingleSelector = func(title string, items []config.ModelItem) (string, error) {
+		tuiItems := make([]tui.SelectItem, len(items))
+		for i, item := range items {
+			tuiItems[i] = tui.SelectItem{Name: item.Name, Description: item.Description, Recommended: item.Recommended}
+		}
+		result, err := tui.SelectSingle(title, tuiItems)
+		if errors.Is(err, tui.ErrCancelled) {
+			return "", config.ErrCancelled
+		}
+		return result, err
+	}
+
+	config.DefaultMultiSelector = func(title string, items []config.ModelItem, preChecked []string) ([]string, error) {
+		tuiItems := make([]tui.SelectItem, len(items))
+		for i, item := range items {
+			tuiItems[i] = tui.SelectItem{Name: item.Name, Description: item.Description, Recommended: item.Recommended}
+		}
+		result, err := tui.SelectMultiple(title, tuiItems, preChecked)
+		if errors.Is(err, tui.ErrCancelled) {
+			return nil, config.ErrCancelled
+		}
+		return result, err
+	}
+
+	config.DefaultSignIn = func(modelName, signInURL string) (string, error) {
+		userName, err := tui.RunSignIn(modelName, signInURL)
+		if errors.Is(err, tui.ErrCancelled) {
+			return "", config.ErrCancelled
+		}
+		return userName, err
+	}
+
+	config.DefaultConfirmPrompt = func(prompt string) (bool, error) {
+		ok, err := tui.RunConfirm(prompt)
+		if errors.Is(err, tui.ErrCancelled) {
+			return false, config.ErrCancelled
+		}
+		return ok, err
+	}
+}
+
+const ConnectInstructions = "If your browser did not open, navigate to:\n    %s\n\n"
 
 // ensureThinkingSupport emits a warning if the model does not advertise thinking support
 func ensureThinkingSupport(ctx context.Context, client *api.Client, name string) {
@@ -365,14 +412,25 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 		return err
 	} else if info.RemoteHost != "" {
 		// Cloud model, no need to load/unload
+
+		isCloud := strings.HasPrefix(info.RemoteHost, "https://ollama.com")
+
+		// Check if user is signed in for ollama.com cloud models
+		if isCloud {
+			if _, err := client.Whoami(cmd.Context()); err != nil {
+				return err
+			}
+		}
+
 		if opts.ShowConnect {
 			p.StopAndClear()
-			if strings.HasPrefix(info.RemoteHost, "https://ollama.com") {
+			if isCloud {
 				fmt.Fprintf(os.Stderr, "Connecting to '%s' on 'ollama.com' ⚡\n", info.RemoteModel)
 			} else {
 				fmt.Fprintf(os.Stderr, "Connecting to '%s' on '%s'\n", info.RemoteModel, info.RemoteHost)
 			}
 		}
+
 		return nil
 	}
 
@@ -662,6 +720,7 @@ func SigninHandler(cmd *cobra.Command, args []string) error {
 			fmt.Println()
 
 			if aErr.SigninURL != "" {
+				_ = browser.OpenURL(aErr.SigninURL)
 				fmt.Printf(ConnectInstructions, aErr.SigninURL)
 			}
 			return nil
@@ -899,11 +958,11 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 	for _, arg := range args {
 		// Unload the model if it's running before deletion
 		if err := loadOrUnloadModel(cmd, &runOptions{
-			Model:     args[0],
+			Model:     arg,
 			KeepAlive: &api.Duration{Duration: 0},
 		}); err != nil {
 			if !strings.Contains(strings.ToLower(err.Error()), "not found") {
-				fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", args[0])
+				fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", arg)
 			}
 		}
 
@@ -1018,8 +1077,10 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 		}
 
 		if resp.ModelInfo != nil {
-			arch := resp.ModelInfo["general.architecture"].(string)
-			rows = append(rows, []string{"", "architecture", arch})
+			arch, _ := resp.ModelInfo["general.architecture"].(string)
+			if arch != "" {
+				rows = append(rows, []string{"", "architecture", arch})
+			}
 
 			var paramStr string
 			if resp.Details.ParameterSize != "" {
@@ -1029,7 +1090,9 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 					paramStr = format.HumanNumber(uint64(f))
 				}
 			}
-			rows = append(rows, []string{"", "parameters", paramStr})
+			if paramStr != "" {
+				rows = append(rows, []string{"", "parameters", paramStr})
+			}
 
 			if v, ok := resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)]; ok {
 				if f, ok := v.(float64); ok {
@@ -1745,7 +1808,7 @@ func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 			return err
 		}
 		if err := startApp(cmd.Context(), client); err != nil {
-			return fmt.Errorf("ollama server not responding - %w", err)
+			return err
 		}
 	}
 	return nil
@@ -1786,6 +1849,197 @@ Environment Variables:
 	cmd.SetUsageTemplate(cmd.UsageTemplate() + envUsage)
 }
 
+// ensureServerRunning checks if the ollama server is running and starts it in the background if not.
+func ensureServerRunning(ctx context.Context) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	// Check if server is already running
+	if err := client.Heartbeat(ctx); err == nil {
+		return nil // server is already running
+	}
+
+	// Server not running, start it in the background
+	exe, err := os.Executable()
+	if err != nil {
+		return fmt.Errorf("could not find executable: %w", err)
+	}
+
+	serverCmd := exec.CommandContext(ctx, exe, "serve")
+	serverCmd.Env = os.Environ()
+	serverCmd.SysProcAttr = backgroundServerSysProcAttr()
+	if err := serverCmd.Start(); err != nil {
+		return fmt.Errorf("failed to start server: %w", err)
+	}
+
+	// Wait for the server to be ready
+	for {
+		time.Sleep(500 * time.Millisecond)
+		if err := client.Heartbeat(ctx); err == nil {
+			return nil // server has started
+		}
+	}
+}
+
+// runInteractiveTUI runs the main interactive TUI menu.
+func runInteractiveTUI(cmd *cobra.Command) {
+	// Ensure the server is running before showing the TUI
+	if err := ensureServerRunning(cmd.Context()); err != nil {
+		fmt.Fprintf(os.Stderr, "Error starting server: %v\n", err)
+		return
+	}
+
+	// Selector adapters for tui
+	singleSelector := func(title string, items []config.ModelItem) (string, error) {
+		tuiItems := make([]tui.SelectItem, len(items))
+		for i, item := range items {
+			tuiItems[i] = tui.SelectItem{Name: item.Name, Description: item.Description, Recommended: item.Recommended}
+		}
+		result, err := tui.SelectSingle(title, tuiItems)
+		if errors.Is(err, tui.ErrCancelled) {
+			return "", config.ErrCancelled
+		}
+		return result, err
+	}
+
+	multiSelector := func(title string, items []config.ModelItem, preChecked []string) ([]string, error) {
+		tuiItems := make([]tui.SelectItem, len(items))
+		for i, item := range items {
+			tuiItems[i] = tui.SelectItem{Name: item.Name, Description: item.Description, Recommended: item.Recommended}
+		}
+		result, err := tui.SelectMultiple(title, tuiItems, preChecked)
+		if errors.Is(err, tui.ErrCancelled) {
+			return nil, config.ErrCancelled
+		}
+		return result, err
+	}
+
+	for {
+		result, err := tui.Run()
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "Error: %v\n", err)
+			return
+		}
+
+		runModel := func(modelName string) {
+			client, err := api.ClientFromEnvironment()
+			if err != nil {
+				fmt.Fprintf(os.Stderr, "Error: %v\n", err)
+				return
+			}
+			if err := config.ShowOrPull(cmd.Context(), client, modelName); err != nil {
+				if errors.Is(err, config.ErrCancelled) {
+					return
+				}
+				fmt.Fprintf(os.Stderr, "Error: %v\n", err)
+				return
+			}
+			_ = config.SetLastModel(modelName)
+			opts := runOptions{
+				Model:       modelName,
+				WordWrap:    os.Getenv("TERM") == "xterm-256color",
+				Options:     map[string]any{},
+				ShowConnect: true,
+			}
+			if err := loadOrUnloadModel(cmd, &opts); err != nil {
+				fmt.Fprintf(os.Stderr, "Error loading model: %v\n", err)
+				return
+			}
+			if err := generateInteractive(cmd, opts); err != nil {
+				fmt.Fprintf(os.Stderr, "Error running model: %v\n", err)
+			}
+		}
+
+		launchIntegration := func(name string) bool {
+			// If not configured or model no longer exists, prompt for model selection
+			configuredModel := config.IntegrationModel(name)
+			if configuredModel == "" || !config.ModelExists(cmd.Context(), configuredModel) {
+				err := config.ConfigureIntegrationWithSelectors(cmd.Context(), name, singleSelector, multiSelector)
+				if errors.Is(err, config.ErrCancelled) {
+					return false // Return to main menu
+				}
+				if err != nil {
+					fmt.Fprintf(os.Stderr, "Error configuring %s: %v\n", name, err)
+					return true
+				}
+			}
+			if err := config.LaunchIntegration(name); err != nil {
+				fmt.Fprintf(os.Stderr, "Error launching %s: %v\n", name, err)
+			}
+			return true
+		}
+
+		switch result.Selection {
+		case tui.SelectionNone:
+			// User quit
+			return
+		case tui.SelectionRunModel:
+			_ = config.SetLastSelection("run")
+			if modelName := config.LastModel(); modelName != "" {
+				runModel(modelName)
+			} else {
+				modelName, err := config.SelectModelWithSelector(cmd.Context(), singleSelector)
+				if errors.Is(err, config.ErrCancelled) {
+					continue // Return to main menu
+				}
+				if err != nil {
+					fmt.Fprintf(os.Stderr, "Error selecting model: %v\n", err)
+					continue
+				}
+				runModel(modelName)
+			}
+		case tui.SelectionChangeRunModel:
+			_ = config.SetLastSelection("run")
+			// Use model from modal if selected, otherwise show picker
+			modelName := result.Model
+			if modelName == "" {
+				var err error
+				modelName, err = config.SelectModelWithSelector(cmd.Context(), singleSelector)
+				if errors.Is(err, config.ErrCancelled) {
+					continue // Return to main menu
+				}
+				if err != nil {
+					fmt.Fprintf(os.Stderr, "Error selecting model: %v\n", err)
+					continue
+				}
+			}
+			runModel(modelName)
+		case tui.SelectionIntegration:
+			_ = config.SetLastSelection(result.Integration)
+			if !launchIntegration(result.Integration) {
+				continue // Return to main menu
+			}
+		case tui.SelectionChangeIntegration:
+			_ = config.SetLastSelection(result.Integration)
+			// Use model from modal if selected, otherwise show picker
+			if result.Model != "" {
+				// Model already selected from modal - save and launch
+				if err := config.SaveIntegrationModel(result.Integration, result.Model); err != nil {
+					fmt.Fprintf(os.Stderr, "Error saving config: %v\n", err)
+					continue
+				}
+				if err := config.LaunchIntegrationWithModel(result.Integration, result.Model); err != nil {
+					fmt.Fprintf(os.Stderr, "Error launching %s: %v\n", result.Integration, err)
+				}
+			} else {
+				err := config.ConfigureIntegrationWithSelectors(cmd.Context(), result.Integration, singleSelector, multiSelector)
+				if errors.Is(err, config.ErrCancelled) {
+					continue // Return to main menu
+				}
+				if err != nil {
+					fmt.Fprintf(os.Stderr, "Error configuring %s: %v\n", result.Integration, err)
+					continue
+				}
+				if err := config.LaunchIntegration(result.Integration); err != nil {
+					fmt.Fprintf(os.Stderr, "Error launching %s: %v\n", result.Integration, err)
+				}
+			}
+		}
+	}
+}
+
 func NewCLI() *cobra.Command {
 	log.SetFlags(log.LstdFlags | log.Lshortfile)
 	cobra.EnableCommandSorting = false
@@ -1808,11 +2062,13 @@ func NewCLI() *cobra.Command {
 				return
 			}
 
-			cmd.Print(cmd.UsageString())
+			runInteractiveTUI(cmd)
 		},
 	}
 
 	rootCmd.Flags().BoolP("version", "v", false, "Show version information")
+	rootCmd.Flags().Bool("verbose", false, "Show timings for response")
+	rootCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
 
 	createCmd := &cobra.Command{
 		Use:   "create MODEL",
@@ -1883,7 +2139,7 @@ func NewCLI() *cobra.Command {
 	serveCmd := &cobra.Command{
 		Use:     "serve",
 		Aliases: []string{"start"},
-		Short:   "Start ollama",
+		Short:   "Start Ollama",
 		Args:    cobra.ExactArgs(0),
 		RunE:    RunServer,
 	}
@@ -1916,6 +2172,15 @@ func NewCLI() *cobra.Command {
 		RunE:    SigninHandler,
 	}
 
+	loginCmd := &cobra.Command{
+		Use:     "login",
+		Short:   "Sign in to ollama.com",
+		Hidden:  true,
+		Args:    cobra.ExactArgs(0),
+		PreRunE: checkServerHeartbeat,
+		RunE:    SigninHandler,
+	}
+
 	signoutCmd := &cobra.Command{
 		Use:     "signout",
 		Short:   "Sign out from ollama.com",
@@ -1924,6 +2189,15 @@ func NewCLI() *cobra.Command {
 		RunE:    SignoutHandler,
 	}
 
+	logoutCmd := &cobra.Command{
+		Use:     "logout",
+		Short:   "Sign out from ollama.com",
+		Hidden:  true,
+		Args:    cobra.ExactArgs(0),
+		PreRunE: checkServerHeartbeat,
+		RunE:    SignoutHandler,
+	}
+
 	listCmd := &cobra.Command{
 		Use:     "list",
 		Aliases: []string{"ls"},
@@ -2020,12 +2294,15 @@ func NewCLI() *cobra.Command {
 		pullCmd,
 		pushCmd,
 		signinCmd,
+		loginCmd,
 		signoutCmd,
+		logoutCmd,
 		listCmd,
 		psCmd,
 		copyCmd,
 		deleteCmd,
 		runnerCmd,
+		config.LaunchCmd(checkServerHeartbeat, runInteractiveTUI),
 	)
 
 	return rootCmd

cmd/cmd_test.go

@@ -3,6 +3,7 @@ package cmd
 import (
 	"bytes"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
@@ -1553,7 +1554,7 @@ func TestShowInfoImageGen(t *testing.T) {
 		Details: api.ModelDetails{
 			Family:            "ZImagePipeline",
 			ParameterSize:     "10.3B",
-			QuantizationLevel: "FP8",
+			QuantizationLevel: "Q8",
 		},
 		Capabilities: []model.Capability{model.CapabilityImage},
 		Requires:     "0.14.0",
@@ -1565,7 +1566,7 @@ func TestShowInfoImageGen(t *testing.T) {
 	expect := "  Model\n" +
 		"    architecture    ZImagePipeline    \n" +
 		"    parameters      10.3B             \n" +
-		"    quantization    FP8               \n" +
+		"    quantization    Q8                \n" +
 		"    requires        0.14.0            \n" +
 		"\n" +
 		"  Capabilities\n" +
@@ -1659,3 +1660,103 @@ func TestRunOptions_Copy_Independence(t *testing.T) {
 		t.Error("Copy Think should not be affected by original modification")
 	}
 }
+
+func TestLoadOrUnloadModel_CloudModelAuth(t *testing.T) {
+	tests := []struct {
+		name          string
+		remoteHost    string
+		whoamiStatus  int
+		whoamiResp    any
+		expectedError string
+	}{
+		{
+			name:         "ollama.com cloud model - user signed in",
+			remoteHost:   "https://ollama.com",
+			whoamiStatus: http.StatusOK,
+			whoamiResp:   api.UserResponse{Name: "testuser"},
+		},
+		{
+			name:         "ollama.com cloud model - user not signed in",
+			remoteHost:   "https://ollama.com",
+			whoamiStatus: http.StatusUnauthorized,
+			whoamiResp: map[string]string{
+				"error":      "unauthorized",
+				"signin_url": "https://ollama.com/signin",
+			},
+			expectedError: "unauthorized",
+		},
+		{
+			name:         "non-ollama.com remote - no auth check",
+			remoteHost:   "https://other-remote.com",
+			whoamiStatus: http.StatusUnauthorized, // should not be called
+			whoamiResp:   nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			whoamiCalled := false
+			mockServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				switch r.URL.Path {
+				case "/api/show":
+					w.Header().Set("Content-Type", "application/json")
+					if err := json.NewEncoder(w).Encode(api.ShowResponse{
+						RemoteHost:  tt.remoteHost,
+						RemoteModel: "test-model",
+					}); err != nil {
+						http.Error(w, err.Error(), http.StatusInternalServerError)
+					}
+				case "/api/me":
+					whoamiCalled = true
+					w.Header().Set("Content-Type", "application/json")
+					w.WriteHeader(tt.whoamiStatus)
+					if tt.whoamiResp != nil {
+						if err := json.NewEncoder(w).Encode(tt.whoamiResp); err != nil {
+							http.Error(w, err.Error(), http.StatusInternalServerError)
+						}
+					}
+				default:
+					http.NotFound(w, r)
+				}
+			}))
+			defer mockServer.Close()
+
+			t.Setenv("OLLAMA_HOST", mockServer.URL)
+
+			cmd := &cobra.Command{}
+			cmd.SetContext(t.Context())
+
+			opts := &runOptions{
+				Model:       "test-cloud-model",
+				ShowConnect: false,
+			}
+
+			err := loadOrUnloadModel(cmd, opts)
+
+			if strings.HasPrefix(tt.remoteHost, "https://ollama.com") {
+				if !whoamiCalled {
+					t.Error("expected whoami to be called for ollama.com cloud model")
+				}
+			} else {
+				if whoamiCalled {
+					t.Error("whoami should not be called for non-ollama.com remote")
+				}
+			}
+
+			if tt.expectedError != "" {
+				if err == nil {
+					t.Errorf("expected error containing %q, got nil", tt.expectedError)
+				} else {
+					var authErr api.AuthorizationError
+					if !errors.As(err, &authErr) {
+						t.Errorf("expected AuthorizationError, got %T: %v", err, err)
+					}
+				}
+			} else {
+				if err != nil {
+					t.Errorf("expected no error, got %v", err)
+				}
+			}
+		})
+	}
+}

cmd/config/claude.go

@@ -0,0 +1,192 @@
+package config
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
+)
+
+// Claude implements Runner and AliasConfigurer for Claude Code integration
+type Claude struct{}
+
+// Compile-time check that Claude implements AliasConfigurer
+var _ AliasConfigurer = (*Claude)(nil)
+
+func (c *Claude) String() string { return "Claude Code" }
+
+func (c *Claude) args(model string, extra []string) []string {
+	var args []string
+	if model != "" {
+		args = append(args, "--model", model)
+	}
+	args = append(args, extra...)
+	return args
+}
+
+func (c *Claude) findPath() (string, error) {
+	if p, err := exec.LookPath("claude"); err == nil {
+		return p, nil
+	}
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", err
+	}
+	name := "claude"
+	if runtime.GOOS == "windows" {
+		name = "claude.exe"
+	}
+	fallback := filepath.Join(home, ".claude", "local", name)
+	if _, err := os.Stat(fallback); err != nil {
+		return "", err
+	}
+	return fallback, nil
+}
+
+func (c *Claude) Run(model string, args []string) error {
+	claudePath, err := c.findPath()
+	if err != nil {
+		return fmt.Errorf("claude is not installed, install from https://code.claude.com/docs/en/quickstart")
+	}
+
+	cmd := exec.Command(claudePath, c.args(model, args)...)
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+
+	env := append(os.Environ(),
+		"ANTHROPIC_BASE_URL="+envconfig.Host().String(),
+		"ANTHROPIC_API_KEY=",
+		"ANTHROPIC_AUTH_TOKEN=ollama",
+	)
+
+	env = append(env, c.modelEnvVars(model)...)
+
+	cmd.Env = env
+	return cmd.Run()
+}
+
+// modelEnvVars returns Claude Code env vars that route all model tiers through Ollama.
+func (c *Claude) modelEnvVars(model string) []string {
+	primary := model
+	fast := model
+	if cfg, err := loadIntegration("claude"); err == nil && cfg.Aliases != nil {
+		if p := cfg.Aliases["primary"]; p != "" {
+			primary = p
+		}
+		if f := cfg.Aliases["fast"]; f != "" {
+			fast = f
+		}
+	}
+	return []string{
+		"ANTHROPIC_DEFAULT_OPUS_MODEL=" + primary,
+		"ANTHROPIC_DEFAULT_SONNET_MODEL=" + primary,
+		"ANTHROPIC_DEFAULT_HAIKU_MODEL=" + fast,
+		"CLAUDE_CODE_SUBAGENT_MODEL=" + primary,
+	}
+}
+
+// ConfigureAliases sets up model aliases for Claude Code.
+// model: the model to use (if empty, user will be prompted to select)
+// aliases: existing alias configuration to preserve/update
+// Cloud-only: subagent routing (fast model) is gated to cloud models only until
+// there is a better strategy for prompt caching on local models.
+func (c *Claude) ConfigureAliases(ctx context.Context, model string, existingAliases map[string]string, force bool) (map[string]string, bool, error) {
+	aliases := make(map[string]string)
+	for k, v := range existingAliases {
+		aliases[k] = v
+	}
+
+	if model != "" {
+		aliases["primary"] = model
+	}
+
+	if !force && aliases["primary"] != "" {
+		client, _ := api.ClientFromEnvironment()
+		if isCloudModel(ctx, client, aliases["primary"]) {
+			if isCloudModel(ctx, client, aliases["fast"]) {
+				return aliases, false, nil
+			}
+		} else {
+			delete(aliases, "fast")
+			return aliases, false, nil
+		}
+	}
+
+	items, existingModels, cloudModels, client, err := listModels(ctx)
+	if err != nil {
+		return nil, false, err
+	}
+
+	fmt.Fprintf(os.Stderr, "\n%sModel Configuration%s\n\n", ansiBold, ansiReset)
+
+	if aliases["primary"] == "" || force {
+		primary, err := DefaultSingleSelector("Select model:", items)
+		if err != nil {
+			return nil, false, err
+		}
+		if err := pullIfNeeded(ctx, client, existingModels, primary); err != nil {
+			return nil, false, err
+		}
+		if err := ensureAuth(ctx, client, cloudModels, []string{primary}); err != nil {
+			return nil, false, err
+		}
+		aliases["primary"] = primary
+	}
+
+	if isCloudModel(ctx, client, aliases["primary"]) {
+		if aliases["fast"] == "" || !isCloudModel(ctx, client, aliases["fast"]) {
+			aliases["fast"] = aliases["primary"]
+		}
+	} else {
+		delete(aliases, "fast")
+	}
+
+	return aliases, true, nil
+}
+
+// SetAliases syncs the configured aliases to the Ollama server using prefix matching.
+// Cloud-only: for local models (fast is empty), we delete any existing aliases to
+// prevent stale routing to a previous cloud model.
+func (c *Claude) SetAliases(ctx context.Context, aliases map[string]string) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	prefixes := []string{"claude-sonnet-", "claude-haiku-"}
+
+	if aliases["fast"] == "" {
+		for _, prefix := range prefixes {
+			_ = client.DeleteAliasExperimental(ctx, &api.AliasDeleteRequest{Alias: prefix})
+		}
+		return nil
+	}
+
+	prefixAliases := map[string]string{
+		"claude-sonnet-": aliases["primary"],
+		"claude-haiku-":  aliases["fast"],
+	}
+
+	var errs []string
+	for prefix, target := range prefixAliases {
+		req := &api.AliasRequest{
+			Alias:          prefix,
+			Target:         target,
+			PrefixMatching: true,
+		}
+		if err := client.SetAliasExperimental(ctx, req); err != nil {
+			errs = append(errs, prefix)
+		}
+	}
+
+	if len(errs) > 0 {
+		return fmt.Errorf("failed to set aliases: %v", errs)
+	}
+	return nil
+}

cmd/config/claude_test.go

@@ -0,0 +1,198 @@
+package config
+
+import (
+	"os"
+	"path/filepath"
+	"runtime"
+	"slices"
+	"strings"
+	"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 TestClaudeFindPath(t *testing.T) {
+	c := &Claude{}
+
+	t.Run("finds claude in PATH", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		name := "claude"
+		if runtime.GOOS == "windows" {
+			name = "claude.exe"
+		}
+		fakeBin := filepath.Join(tmpDir, name)
+		os.WriteFile(fakeBin, []byte("#!/bin/sh\n"), 0o755)
+		t.Setenv("PATH", tmpDir)
+
+		got, err := c.findPath()
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if got != fakeBin {
+			t.Errorf("findPath() = %q, want %q", got, fakeBin)
+		}
+	})
+
+	t.Run("falls back to ~/.claude/local/claude", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		t.Setenv("PATH", t.TempDir()) // empty dir, no claude binary
+
+		name := "claude"
+		if runtime.GOOS == "windows" {
+			name = "claude.exe"
+		}
+		fallback := filepath.Join(tmpDir, ".claude", "local", name)
+		os.MkdirAll(filepath.Dir(fallback), 0o755)
+		os.WriteFile(fallback, []byte("#!/bin/sh\n"), 0o755)
+
+		got, err := c.findPath()
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if got != fallback {
+			t.Errorf("findPath() = %q, want %q", got, fallback)
+		}
+	})
+
+	t.Run("returns error when neither PATH nor fallback exists", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		t.Setenv("PATH", t.TempDir()) // empty dir, no claude binary
+
+		_, err := c.findPath()
+		if err == nil {
+			t.Fatal("expected error, got nil")
+		}
+	})
+}
+
+func TestClaudeArgs(t *testing.T) {
+	c := &Claude{}
+
+	tests := []struct {
+		name  string
+		model string
+		args  []string
+		want  []string
+	}{
+		{"with model", "llama3.2", nil, []string{"--model", "llama3.2"}},
+		{"empty model", "", nil, nil},
+		{"with model and verbose", "llama3.2", []string{"--verbose"}, []string{"--model", "llama3.2", "--verbose"}},
+		{"empty model with help", "", []string{"--help"}, []string{"--help"}},
+		{"with allowed tools", "llama3.2", []string{"--allowedTools", "Read,Write,Bash"}, []string{"--model", "llama3.2", "--allowedTools", "Read,Write,Bash"}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := c.args(tt.model, tt.args)
+			if !slices.Equal(got, tt.want) {
+				t.Errorf("args(%q, %v) = %v, want %v", tt.model, tt.args, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestClaudeModelEnvVars(t *testing.T) {
+	c := &Claude{}
+
+	envMap := func(envs []string) map[string]string {
+		m := make(map[string]string)
+		for _, e := range envs {
+			k, v, _ := strings.Cut(e, "=")
+			m[k] = v
+		}
+		return m
+	}
+
+	t.Run("falls back to model param when no aliases saved", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		got := envMap(c.modelEnvVars("llama3.2"))
+		if got["ANTHROPIC_DEFAULT_OPUS_MODEL"] != "llama3.2" {
+			t.Errorf("OPUS = %q, want llama3.2", got["ANTHROPIC_DEFAULT_OPUS_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_SONNET_MODEL"] != "llama3.2" {
+			t.Errorf("SONNET = %q, want llama3.2", got["ANTHROPIC_DEFAULT_SONNET_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_HAIKU_MODEL"] != "llama3.2" {
+			t.Errorf("HAIKU = %q, want llama3.2", got["ANTHROPIC_DEFAULT_HAIKU_MODEL"])
+		}
+		if got["CLAUDE_CODE_SUBAGENT_MODEL"] != "llama3.2" {
+			t.Errorf("SUBAGENT = %q, want llama3.2", got["CLAUDE_CODE_SUBAGENT_MODEL"])
+		}
+	})
+
+	t.Run("uses primary alias for opus sonnet and subagent", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		saveIntegration("claude", []string{"qwen3:8b"})
+		saveAliases("claude", map[string]string{"primary": "qwen3:8b"})
+
+		got := envMap(c.modelEnvVars("qwen3:8b"))
+		if got["ANTHROPIC_DEFAULT_OPUS_MODEL"] != "qwen3:8b" {
+			t.Errorf("OPUS = %q, want qwen3:8b", got["ANTHROPIC_DEFAULT_OPUS_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_SONNET_MODEL"] != "qwen3:8b" {
+			t.Errorf("SONNET = %q, want qwen3:8b", got["ANTHROPIC_DEFAULT_SONNET_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_HAIKU_MODEL"] != "qwen3:8b" {
+			t.Errorf("HAIKU = %q, want qwen3:8b (no fast alias)", got["ANTHROPIC_DEFAULT_HAIKU_MODEL"])
+		}
+		if got["CLAUDE_CODE_SUBAGENT_MODEL"] != "qwen3:8b" {
+			t.Errorf("SUBAGENT = %q, want qwen3:8b", got["CLAUDE_CODE_SUBAGENT_MODEL"])
+		}
+	})
+
+	t.Run("uses fast alias for haiku", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		saveIntegration("claude", []string{"llama3.2:70b"})
+		saveAliases("claude", map[string]string{
+			"primary": "llama3.2:70b",
+			"fast":    "llama3.2:8b",
+		})
+
+		got := envMap(c.modelEnvVars("llama3.2:70b"))
+		if got["ANTHROPIC_DEFAULT_OPUS_MODEL"] != "llama3.2:70b" {
+			t.Errorf("OPUS = %q, want llama3.2:70b", got["ANTHROPIC_DEFAULT_OPUS_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_SONNET_MODEL"] != "llama3.2:70b" {
+			t.Errorf("SONNET = %q, want llama3.2:70b", got["ANTHROPIC_DEFAULT_SONNET_MODEL"])
+		}
+		if got["ANTHROPIC_DEFAULT_HAIKU_MODEL"] != "llama3.2:8b" {
+			t.Errorf("HAIKU = %q, want llama3.2:8b", got["ANTHROPIC_DEFAULT_HAIKU_MODEL"])
+		}
+		if got["CLAUDE_CODE_SUBAGENT_MODEL"] != "llama3.2:70b" {
+			t.Errorf("SUBAGENT = %q, want llama3.2:70b", got["CLAUDE_CODE_SUBAGENT_MODEL"])
+		}
+	})
+
+	t.Run("alias primary overrides model param", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		saveIntegration("claude", []string{"saved-model"})
+		saveAliases("claude", map[string]string{"primary": "saved-model"})
+
+		got := envMap(c.modelEnvVars("different-model"))
+		if got["ANTHROPIC_DEFAULT_OPUS_MODEL"] != "saved-model" {
+			t.Errorf("OPUS = %q, want saved-model", got["ANTHROPIC_DEFAULT_OPUS_MODEL"])
+		}
+	})
+}

cmd/config/codex.go

@@ -0,0 +1,62 @@
+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, extra []string) []string {
+	args := []string{"--oss"}
+	if model != "" {
+		args = append(args, "-m", model)
+	}
+	args = append(args, extra...)
+	return args
+}
+
+func (c *Codex) Run(model string, args []string) error {
+	if err := checkCodexVersion(); err != nil {
+		return err
+	}
+
+	cmd := exec.Command("codex", c.args(model, args)...)
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	return cmd.Run()
+}
+
+func checkCodexVersion() error {
+	if _, err := exec.LookPath("codex"); err != nil {
+		return fmt.Errorf("codex is not installed, install with: npm install -g @openai/codex")
+	}
+
+	out, err := exec.Command("codex", "--version").Output()
+	if err != nil {
+		return fmt.Errorf("failed to get codex version: %w", err)
+	}
+
+	// Parse output like "codex-cli 0.87.0"
+	fields := strings.Fields(strings.TrimSpace(string(out)))
+	if len(fields) < 2 {
+		return fmt.Errorf("unexpected codex version output: %s", string(out))
+	}
+
+	version := "v" + fields[len(fields)-1]
+	minVersion := "v0.81.0"
+
+	if semver.Compare(version, minVersion) < 0 {
+		return fmt.Errorf("codex version %s is too old, minimum required is %s, update with: npm update -g @openai/codex", fields[len(fields)-1], "0.81.0")
+	}
+
+	return nil
+}

cmd/config/codex_test.go

@@ -0,0 +1,31 @@
+package config
+
+import (
+	"slices"
+	"testing"
+)
+
+func TestCodexArgs(t *testing.T) {
+	c := &Codex{}
+
+	tests := []struct {
+		name  string
+		model string
+		args  []string
+		want  []string
+	}{
+		{"with model", "llama3.2", nil, []string{"--oss", "-m", "llama3.2"}},
+		{"empty model", "", nil, []string{"--oss"}},
+		{"with model and profile", "qwen3-coder", []string{"-p", "myprofile"}, []string{"--oss", "-m", "qwen3-coder", "-p", "myprofile"}},
+		{"with sandbox flag", "llama3.2", []string{"--sandbox", "workspace-write"}, []string{"--oss", "-m", "llama3.2", "--sandbox", "workspace-write"}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := c.args(tt.model, tt.args)
+			if !slices.Equal(got, tt.want) {
+				t.Errorf("args(%q, %v) = %v, want %v", tt.model, tt.args, got, tt.want)
+			}
+		})
+	}
+}

cmd/config/config.go

@@ -0,0 +1,271 @@
+// Package config provides integration configuration for external coding tools
+// (Claude Code, Codex, Droid, OpenCode) to use Ollama models.
+package config
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+)
+
+type integration struct {
+	Models  []string          `json:"models"`
+	Aliases map[string]string `json:"aliases,omitempty"`
+}
+
+type config struct {
+	Integrations  map[string]*integration `json:"integrations"`
+	LastModel     string                  `json:"last_model,omitempty"`
+	LastSelection string                  `json:"last_selection,omitempty"` // "run" or integration name
+}
+
+func configPath() (string, error) {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(home, ".ollama", "config.json"), nil
+}
+
+func legacyConfigPath() (string, error) {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(home, ".ollama", "config", "config.json"), nil
+}
+
+// migrateConfig moves the config from the legacy path to ~/.ollama/config.json
+func migrateConfig() (bool, error) {
+	oldPath, err := legacyConfigPath()
+	if err != nil {
+		return false, err
+	}
+
+	oldData, err := os.ReadFile(oldPath)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return false, nil
+		}
+		return false, err
+	}
+
+	var js json.RawMessage
+	if err := json.Unmarshal(oldData, &js); err != nil {
+		return false, nil
+	}
+
+	newPath, err := configPath()
+	if err != nil {
+		return false, err
+	}
+
+	if err := os.MkdirAll(filepath.Dir(newPath), 0o755); err != nil {
+		return false, err
+	}
+	if err := os.WriteFile(newPath, oldData, 0o644); err != nil {
+		return false, fmt.Errorf("write new config: %w", err)
+	}
+
+	_ = os.Remove(oldPath)
+	_ = os.Remove(filepath.Dir(oldPath)) // clean up empty directory
+
+	return true, nil
+}
+
+func load() (*config, error) {
+	path, err := configPath()
+	if err != nil {
+		return nil, err
+	}
+
+	data, err := os.ReadFile(path)
+	if err != nil && os.IsNotExist(err) {
+		if migrated, merr := migrateConfig(); merr == nil && migrated {
+			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
+	}
+
+	key := strings.ToLower(appName)
+	existing := cfg.Integrations[key]
+	var aliases map[string]string
+	if existing != nil && existing.Aliases != nil {
+		aliases = existing.Aliases
+	}
+
+	cfg.Integrations[key] = &integration{
+		Models:  models,
+		Aliases: aliases,
+	}
+
+	return save(cfg)
+}
+
+// IntegrationModel returns the first configured model for an integration, or empty string if not configured.
+func IntegrationModel(appName string) string {
+	ic, err := loadIntegration(appName)
+	if err != nil || len(ic.Models) == 0 {
+		return ""
+	}
+	return ic.Models[0]
+}
+
+// LastModel returns the last model that was run, or empty string if none.
+func LastModel() string {
+	cfg, err := load()
+	if err != nil {
+		return ""
+	}
+	return cfg.LastModel
+}
+
+// SetLastModel saves the last model that was run.
+func SetLastModel(model string) error {
+	cfg, err := load()
+	if err != nil {
+		return err
+	}
+	cfg.LastModel = model
+	return save(cfg)
+}
+
+// LastSelection returns the last menu selection ("run" or integration name), or empty string if none.
+func LastSelection() string {
+	cfg, err := load()
+	if err != nil {
+		return ""
+	}
+	return cfg.LastSelection
+}
+
+// SetLastSelection saves the last menu selection ("run" or integration name).
+func SetLastSelection(selection string) error {
+	cfg, err := load()
+	if err != nil {
+		return err
+	}
+	cfg.LastSelection = selection
+	return save(cfg)
+}
+
+// ModelExists checks if a model exists on the Ollama server.
+func ModelExists(ctx context.Context, name string) bool {
+	if name == "" {
+		return false
+	}
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return false
+	}
+	models, err := client.List(ctx)
+	if err != nil {
+		return false
+	}
+	for _, m := range models.Models {
+		if m.Name == name || strings.HasPrefix(m.Name, name+":") {
+			return true
+		}
+	}
+	return false
+}
+
+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 saveAliases(appName string, aliases map[string]string) error {
+	if appName == "" {
+		return errors.New("app name cannot be empty")
+	}
+
+	cfg, err := load()
+	if err != nil {
+		return err
+	}
+
+	key := strings.ToLower(appName)
+	existing := cfg.Integrations[key]
+	if existing == nil {
+		existing = &integration{}
+	}
+
+	// Replace aliases entirely (not merge) so deletions are persisted
+	existing.Aliases = aliases
+
+	cfg.Integrations[key] = existing
+	return save(cfg)
+}
+
+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_cloud_test.go

@@ -0,0 +1,677 @@
+package config
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestSetAliases_CloudModel(t *testing.T) {
+	// Test the SetAliases logic by checking the alias map behavior
+	aliases := map[string]string{
+		"primary": "kimi-k2.5:cloud",
+		"fast":    "kimi-k2.5:cloud",
+	}
+
+	// Verify fast is set (cloud model behavior)
+	if aliases["fast"] == "" {
+		t.Error("cloud model should have fast alias set")
+	}
+	if aliases["fast"] != aliases["primary"] {
+		t.Errorf("fast should equal primary for auto-set, got fast=%q primary=%q", aliases["fast"], aliases["primary"])
+	}
+}
+
+func TestSetAliases_LocalModel(t *testing.T) {
+	aliases := map[string]string{
+		"primary": "llama3.2:latest",
+	}
+	// Simulate local model behavior: fast should be empty
+	delete(aliases, "fast")
+
+	if aliases["fast"] != "" {
+		t.Error("local model should have empty fast alias")
+	}
+}
+
+func TestSaveAliases_ReplacesNotMerges(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// First save with both primary and fast
+	initial := map[string]string{
+		"primary": "cloud-model",
+		"fast":    "cloud-model",
+	}
+	if err := saveAliases("claude", initial); err != nil {
+		t.Fatalf("failed to save initial aliases: %v", err)
+	}
+
+	// Verify both are saved
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if loaded.Aliases["fast"] != "cloud-model" {
+		t.Errorf("expected fast=cloud-model, got %q", loaded.Aliases["fast"])
+	}
+
+	// Now save without fast (simulating switch to local model)
+	updated := map[string]string{
+		"primary": "local-model",
+		// fast intentionally missing
+	}
+	if err := saveAliases("claude", updated); err != nil {
+		t.Fatalf("failed to save updated aliases: %v", err)
+	}
+
+	// Verify fast is GONE (not merged/preserved)
+	loaded, err = loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load after update: %v", err)
+	}
+	if loaded.Aliases["fast"] != "" {
+		t.Errorf("fast should be removed after saving without it, got %q", loaded.Aliases["fast"])
+	}
+	if loaded.Aliases["primary"] != "local-model" {
+		t.Errorf("primary should be updated to local-model, got %q", loaded.Aliases["primary"])
+	}
+}
+
+func TestSaveAliases_PreservesModels(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// First save integration with models
+	if err := saveIntegration("claude", []string{"model1", "model2"}); err != nil {
+		t.Fatalf("failed to save integration: %v", err)
+	}
+
+	// Then update aliases
+	aliases := map[string]string{"primary": "new-model"}
+	if err := saveAliases("claude", aliases); err != nil {
+		t.Fatalf("failed to save aliases: %v", err)
+	}
+
+	// Verify models are preserved
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if len(loaded.Models) != 2 || loaded.Models[0] != "model1" {
+		t.Errorf("models should be preserved, got %v", loaded.Models)
+	}
+}
+
+// TestSaveAliases_EmptyMap clears all aliases
+func TestSaveAliases_EmptyMap(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Save with aliases
+	if err := saveAliases("claude", map[string]string{"primary": "model", "fast": "model"}); err != nil {
+		t.Fatalf("failed to save: %v", err)
+	}
+
+	// Save empty map
+	if err := saveAliases("claude", map[string]string{}); err != nil {
+		t.Fatalf("failed to save empty: %v", err)
+	}
+
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if len(loaded.Aliases) != 0 {
+		t.Errorf("aliases should be empty, got %v", loaded.Aliases)
+	}
+}
+
+// TestSaveAliases_NilMap handles nil gracefully
+func TestSaveAliases_NilMap(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Save with aliases first
+	if err := saveAliases("claude", map[string]string{"primary": "model"}); err != nil {
+		t.Fatalf("failed to save: %v", err)
+	}
+
+	// Save nil map - should clear aliases
+	if err := saveAliases("claude", nil); err != nil {
+		t.Fatalf("failed to save nil: %v", err)
+	}
+
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if len(loaded.Aliases) > 0 {
+		t.Errorf("aliases should be nil or empty, got %v", loaded.Aliases)
+	}
+}
+
+// TestSaveAliases_EmptyAppName returns error
+func TestSaveAliases_EmptyAppName(t *testing.T) {
+	err := saveAliases("", map[string]string{"primary": "model"})
+	if err == nil {
+		t.Error("expected error for empty app name")
+	}
+}
+
+func TestSaveAliases_CaseInsensitive(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	if err := saveAliases("Claude", map[string]string{"primary": "model1"}); err != nil {
+		t.Fatalf("failed to save: %v", err)
+	}
+
+	// Load with different case
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if loaded.Aliases["primary"] != "model1" {
+		t.Errorf("expected primary=model1, got %q", loaded.Aliases["primary"])
+	}
+
+	// Update with different case
+	if err := saveAliases("CLAUDE", map[string]string{"primary": "model2"}); err != nil {
+		t.Fatalf("failed to update: %v", err)
+	}
+
+	loaded, err = loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load after update: %v", err)
+	}
+	if loaded.Aliases["primary"] != "model2" {
+		t.Errorf("expected primary=model2, got %q", loaded.Aliases["primary"])
+	}
+}
+
+// TestSaveAliases_CreatesIntegration creates integration if it doesn't exist
+func TestSaveAliases_CreatesIntegration(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Save aliases for non-existent integration
+	if err := saveAliases("newintegration", map[string]string{"primary": "model"}); err != nil {
+		t.Fatalf("failed to save: %v", err)
+	}
+
+	loaded, err := loadIntegration("newintegration")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if loaded.Aliases["primary"] != "model" {
+		t.Errorf("expected primary=model, got %q", loaded.Aliases["primary"])
+	}
+}
+
+func TestConfigureAliases_AliasMap(t *testing.T) {
+	t.Run("cloud model auto-sets fast to primary", func(t *testing.T) {
+		aliases := make(map[string]string)
+		aliases["primary"] = "cloud-model"
+
+		// Simulate cloud model behavior
+		isCloud := true
+		if isCloud {
+			if aliases["fast"] == "" {
+				aliases["fast"] = aliases["primary"]
+			}
+		}
+
+		if aliases["fast"] != "cloud-model" {
+			t.Errorf("expected fast=cloud-model, got %q", aliases["fast"])
+		}
+	})
+
+	t.Run("cloud model preserves custom fast", func(t *testing.T) {
+		aliases := map[string]string{
+			"primary": "cloud-model",
+			"fast":    "custom-fast-model",
+		}
+
+		// Simulate cloud model behavior - should preserve existing fast
+		isCloud := true
+		if isCloud {
+			if aliases["fast"] == "" {
+				aliases["fast"] = aliases["primary"]
+			}
+		}
+
+		if aliases["fast"] != "custom-fast-model" {
+			t.Errorf("expected fast=custom-fast-model (preserved), got %q", aliases["fast"])
+		}
+	})
+
+	t.Run("local model clears fast", func(t *testing.T) {
+		aliases := map[string]string{
+			"primary": "local-model",
+			"fast":    "should-be-cleared",
+		}
+
+		// Simulate local model behavior
+		isCloud := false
+		if !isCloud {
+			delete(aliases, "fast")
+		}
+
+		if aliases["fast"] != "" {
+			t.Errorf("expected fast to be cleared, got %q", aliases["fast"])
+		}
+	})
+
+	t.Run("switching cloud to local clears fast", func(t *testing.T) {
+		// Start with cloud config
+		aliases := map[string]string{
+			"primary": "cloud-model",
+			"fast":    "cloud-model",
+		}
+
+		// Switch to local
+		aliases["primary"] = "local-model"
+		isCloud := false
+		if !isCloud {
+			delete(aliases, "fast")
+		}
+
+		if aliases["fast"] != "" {
+			t.Errorf("fast should be cleared when switching to local, got %q", aliases["fast"])
+		}
+		if aliases["primary"] != "local-model" {
+			t.Errorf("primary should be updated, got %q", aliases["primary"])
+		}
+	})
+
+	t.Run("switching local to cloud sets fast", func(t *testing.T) {
+		// Start with local config (no fast)
+		aliases := map[string]string{
+			"primary": "local-model",
+		}
+
+		// Switch to cloud
+		aliases["primary"] = "cloud-model"
+		isCloud := true
+		if isCloud {
+			if aliases["fast"] == "" {
+				aliases["fast"] = aliases["primary"]
+			}
+		}
+
+		if aliases["fast"] != "cloud-model" {
+			t.Errorf("fast should be set when switching to cloud, got %q", aliases["fast"])
+		}
+	})
+}
+
+func TestSetAliases_PrefixMapping(t *testing.T) {
+	// This tests the expected mapping without needing a real client
+	aliases := map[string]string{
+		"primary": "my-cloud-model",
+		"fast":    "my-fast-model",
+	}
+
+	expectedMappings := map[string]string{
+		"claude-sonnet-": aliases["primary"],
+		"claude-haiku-":  aliases["fast"],
+	}
+
+	if expectedMappings["claude-sonnet-"] != "my-cloud-model" {
+		t.Errorf("claude-sonnet- should map to primary")
+	}
+	if expectedMappings["claude-haiku-"] != "my-fast-model" {
+		t.Errorf("claude-haiku- should map to fast")
+	}
+}
+
+func TestSetAliases_LocalDeletesPrefixes(t *testing.T) {
+	aliases := map[string]string{
+		"primary": "local-model",
+		// fast is empty/missing - indicates local model
+	}
+
+	prefixesToDelete := []string{"claude-sonnet-", "claude-haiku-"}
+
+	// Verify the logic: when fast is empty, we should delete
+	if aliases["fast"] != "" {
+		t.Error("fast should be empty for local model")
+	}
+
+	// Verify we have the right prefixes to delete
+	if len(prefixesToDelete) != 2 {
+		t.Errorf("expected 2 prefixes to delete, got %d", len(prefixesToDelete))
+	}
+}
+
+// TestAtomicUpdate_ServerFailsConfigNotSaved simulates atomic update behavior
+func TestAtomicUpdate_ServerFailsConfigNotSaved(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Simulate: server fails, config should NOT be saved
+	serverErr := errors.New("server unavailable")
+
+	if serverErr == nil {
+		t.Error("config should NOT be saved when server fails")
+	}
+}
+
+// TestAtomicUpdate_ServerSucceedsConfigSaved simulates successful atomic update
+func TestAtomicUpdate_ServerSucceedsConfigSaved(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Simulate: server succeeds, config should be saved
+	var serverErr error
+	if serverErr != nil {
+		t.Fatal("server should succeed")
+	}
+
+	if err := saveAliases("claude", map[string]string{"primary": "model"}); err != nil {
+		t.Fatalf("saveAliases failed: %v", err)
+	}
+
+	// Verify it was actually saved
+	loaded, err := loadIntegration("claude")
+	if err != nil {
+		t.Fatalf("failed to load: %v", err)
+	}
+	if loaded.Aliases["primary"] != "model" {
+		t.Errorf("expected primary=model, got %q", loaded.Aliases["primary"])
+	}
+}
+
+func TestConfigFile_PreservesUnknownFields(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Write config with extra fields
+	configPath := filepath.Join(tmpDir, ".ollama", "config.json")
+	os.MkdirAll(filepath.Dir(configPath), 0o755)
+
+	// Note: Our config struct only has Integrations, so top-level unknown fields
+	// won't be preserved by our current implementation. This test documents that.
+	initialConfig := `{
+  "integrations": {
+    "claude": {
+      "models": ["model1"],
+      "aliases": {"primary": "model1"},
+      "unknownField": "should be lost"
+    }
+  },
+  "topLevelUnknown": "will be lost"
+}`
+	os.WriteFile(configPath, []byte(initialConfig), 0o644)
+
+	// Update aliases
+	if err := saveAliases("claude", map[string]string{"primary": "model2"}); err != nil {
+		t.Fatalf("failed to save: %v", err)
+	}
+
+	// Read raw file to check
+	data, _ := os.ReadFile(configPath)
+	content := string(data)
+
+	// models should be preserved
+	if !contains(content, "model1") {
+		t.Error("models should be preserved")
+	}
+
+	// primary should be updated
+	if !contains(content, "model2") {
+		t.Error("primary should be updated to model2")
+	}
+}
+
+func contains(s, substr string) bool {
+	return len(s) >= len(substr) && (s == substr || len(s) > 0 && containsHelper(s, substr))
+}
+
+func containsHelper(s, substr string) bool {
+	for i := 0; i <= len(s)-len(substr); i++ {
+		if s[i:i+len(substr)] == substr {
+			return true
+		}
+	}
+	return false
+}
+
+func TestClaudeImplementsAliasConfigurer(t *testing.T) {
+	c := &Claude{}
+	var _ AliasConfigurer = c // Compile-time check
+}
+
+func TestModelNameEdgeCases(t *testing.T) {
+	testCases := []struct {
+		name  string
+		model string
+	}{
+		{"simple", "llama3.2"},
+		{"with tag", "llama3.2:latest"},
+		{"with cloud tag", "kimi-k2.5:cloud"},
+		{"with namespace", "library/llama3.2"},
+		{"with dots", "glm-4.7-flash"},
+		{"with numbers", "qwen3:8b"},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			tmpDir := t.TempDir()
+			setTestHome(t, tmpDir)
+
+			aliases := map[string]string{"primary": tc.model}
+			if err := saveAliases("claude", aliases); err != nil {
+				t.Fatalf("failed to save model %q: %v", tc.model, err)
+			}
+
+			loaded, err := loadIntegration("claude")
+			if err != nil {
+				t.Fatalf("failed to load: %v", err)
+			}
+			if loaded.Aliases["primary"] != tc.model {
+				t.Errorf("expected primary=%q, got %q", tc.model, loaded.Aliases["primary"])
+			}
+		})
+	}
+}
+
+func TestSwitchingScenarios(t *testing.T) {
+	t.Run("cloud to local removes fast", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Initial cloud config
+		if err := saveAliases("claude", map[string]string{
+			"primary": "cloud-model",
+			"fast":    "cloud-model",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Switch to local (no fast)
+		if err := saveAliases("claude", map[string]string{
+			"primary": "local-model",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+		if loaded.Aliases["fast"] != "" {
+			t.Errorf("fast should be removed, got %q", loaded.Aliases["fast"])
+		}
+		if loaded.Aliases["primary"] != "local-model" {
+			t.Errorf("primary should be local-model, got %q", loaded.Aliases["primary"])
+		}
+	})
+
+	t.Run("local to cloud adds fast", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Initial local config
+		if err := saveAliases("claude", map[string]string{
+			"primary": "local-model",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Switch to cloud (with fast)
+		if err := saveAliases("claude", map[string]string{
+			"primary": "cloud-model",
+			"fast":    "cloud-model",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+		if loaded.Aliases["fast"] != "cloud-model" {
+			t.Errorf("fast should be cloud-model, got %q", loaded.Aliases["fast"])
+		}
+	})
+
+	t.Run("cloud to different cloud updates both", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Initial cloud config
+		if err := saveAliases("claude", map[string]string{
+			"primary": "cloud-model-1",
+			"fast":    "cloud-model-1",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Switch to different cloud
+		if err := saveAliases("claude", map[string]string{
+			"primary": "cloud-model-2",
+			"fast":    "cloud-model-2",
+		}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+		if loaded.Aliases["primary"] != "cloud-model-2" {
+			t.Errorf("primary should be cloud-model-2, got %q", loaded.Aliases["primary"])
+		}
+		if loaded.Aliases["fast"] != "cloud-model-2" {
+			t.Errorf("fast should be cloud-model-2, got %q", loaded.Aliases["fast"])
+		}
+	})
+}
+
+func TestToolCapabilityFiltering(t *testing.T) {
+	t.Run("all models checked for tool capability", func(t *testing.T) {
+		// Both cloud and local models are checked for tool capability via Show API
+		// Only models with "tools" in capabilities are included
+		m := modelInfo{Name: "tool-model", Remote: false, ToolCapable: true}
+		if !m.ToolCapable {
+			t.Error("tool capable model should be marked as such")
+		}
+	})
+
+	t.Run("modelInfo includes ToolCapable field", func(t *testing.T) {
+		m := modelInfo{Name: "test", Remote: true, ToolCapable: true}
+		if !m.ToolCapable {
+			t.Error("ToolCapable field should be accessible")
+		}
+	})
+}
+
+func TestIsCloudModel_RequiresClient(t *testing.T) {
+	t.Run("nil client always returns false", func(t *testing.T) {
+		// isCloudModel now only uses Show API, no suffix detection
+		if isCloudModel(context.Background(), nil, "model:cloud") {
+			t.Error("nil client should return false regardless of suffix")
+		}
+		if isCloudModel(context.Background(), nil, "local-model") {
+			t.Error("nil client should return false")
+		}
+	})
+}
+
+func TestModelsAndAliasesMustStayInSync(t *testing.T) {
+	t.Run("saveAliases followed by saveIntegration keeps them in sync", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Save aliases with one model
+		if err := saveAliases("claude", map[string]string{"primary": "model-a"}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Save integration with same model (this is the pattern we use)
+		if err := saveIntegration("claude", []string{"model-a"}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+		if loaded.Aliases["primary"] != loaded.Models[0] {
+			t.Errorf("aliases.primary (%q) != models[0] (%q)", loaded.Aliases["primary"], loaded.Models[0])
+		}
+	})
+
+	t.Run("out of sync config is detectable", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Simulate out-of-sync state (like manual edit or bug)
+		if err := saveIntegration("claude", []string{"old-model"}); err != nil {
+			t.Fatal(err)
+		}
+		if err := saveAliases("claude", map[string]string{"primary": "new-model"}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+
+		// They should be different (this is the bug state)
+		if loaded.Models[0] == loaded.Aliases["primary"] {
+			t.Error("expected out-of-sync state for this test")
+		}
+
+		// The fix: when updating aliases, also update models
+		if err := saveIntegration("claude", []string{loaded.Aliases["primary"]}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ = loadIntegration("claude")
+		if loaded.Models[0] != loaded.Aliases["primary"] {
+			t.Errorf("after fix: models[0] (%q) should equal aliases.primary (%q)",
+				loaded.Models[0], loaded.Aliases["primary"])
+		}
+	})
+
+	t.Run("updating primary alias updates models too", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		// Initial state
+		if err := saveIntegration("claude", []string{"initial-model"}); err != nil {
+			t.Fatal(err)
+		}
+		if err := saveAliases("claude", map[string]string{"primary": "initial-model"}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Update aliases AND models together
+		newAliases := map[string]string{"primary": "updated-model"}
+		if err := saveAliases("claude", newAliases); err != nil {
+			t.Fatal(err)
+		}
+		if err := saveIntegration("claude", []string{newAliases["primary"]}); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, _ := loadIntegration("claude")
+		if loaded.Models[0] != "updated-model" {
+			t.Errorf("models[0] should be updated-model, got %q", loaded.Models[0])
+		}
+		if loaded.Aliases["primary"] != "updated-model" {
+			t.Errorf("aliases.primary should be updated-model, got %q", loaded.Aliases["primary"])
+		}
+	})
+}

cmd/config/config_test.go

@@ -0,0 +1,595 @@
+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("save and load aliases", func(t *testing.T) {
+		models := []string{"llama3.2"}
+		if err := saveIntegration("claude", models); err != nil {
+			t.Fatal(err)
+		}
+		aliases := map[string]string{
+			"primary": "llama3.2:70b",
+			"fast":    "llama3.2:8b",
+		}
+		if err := saveAliases("claude", aliases); err != nil {
+			t.Fatal(err)
+		}
+
+		config, err := loadIntegration("claude")
+		if err != nil {
+			t.Fatal(err)
+		}
+		if config.Aliases == nil {
+			t.Fatal("expected aliases to be saved")
+		}
+		for k, v := range aliases {
+			if config.Aliases[k] != v {
+				t.Errorf("alias %s: expected %s, got %s", k, v, config.Aliases[k])
+			}
+		}
+	})
+
+	t.Run("saveIntegration preserves aliases", func(t *testing.T) {
+		if err := saveIntegration("claude", []string{"model-a"}); err != nil {
+			t.Fatal(err)
+		}
+		if err := saveAliases("claude", map[string]string{"primary": "model-a", "fast": "model-small"}); err != nil {
+			t.Fatal(err)
+		}
+
+		if err := saveIntegration("claude", []string{"model-b"}); err != nil {
+			t.Fatal(err)
+		}
+		config, err := loadIntegration("claude")
+		if err != nil {
+			t.Fatal(err)
+		}
+		if config.Aliases["primary"] != "model-a" {
+			t.Errorf("expected aliases to be preserved, got %v", config.Aliases)
+		}
+	})
+
+	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)
+
+	dir := filepath.Join(tmpDir, ".ollama")
+	os.MkdirAll(dir, 0o755)
+	os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{corrupted json`), 0o644)
+
+	_, 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.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 TestMigrateConfig(t *testing.T) {
+	t.Run("migrates legacy file to new location", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		legacyDir := filepath.Join(tmpDir, ".ollama", "config")
+		os.MkdirAll(legacyDir, 0o755)
+		data := []byte(`{"integrations":{"claude":{"models":["llama3.2"]}}}`)
+		os.WriteFile(filepath.Join(legacyDir, "config.json"), data, 0o644)
+
+		migrated, err := migrateConfig()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if !migrated {
+			t.Fatal("expected migration to occur")
+		}
+
+		newPath, _ := configPath()
+		got, err := os.ReadFile(newPath)
+		if err != nil {
+			t.Fatalf("new config not found: %v", err)
+		}
+		if string(got) != string(data) {
+			t.Errorf("content mismatch: got %s", got)
+		}
+
+		if _, err := os.Stat(filepath.Join(legacyDir, "config.json")); !os.IsNotExist(err) {
+			t.Error("legacy file should have been removed")
+		}
+
+		if _, err := os.Stat(legacyDir); !os.IsNotExist(err) {
+			t.Error("legacy directory should have been removed")
+		}
+	})
+
+	t.Run("no-op when no legacy file exists", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		migrated, err := migrateConfig()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if migrated {
+			t.Error("expected no migration")
+		}
+	})
+
+	t.Run("skips corrupt legacy file", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		legacyDir := filepath.Join(tmpDir, ".ollama", "config")
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(legacyDir, "config.json"), []byte(`{corrupt`), 0o644)
+
+		migrated, err := migrateConfig()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if migrated {
+			t.Error("should not migrate corrupt file")
+		}
+
+		if _, err := os.Stat(filepath.Join(legacyDir, "config.json")); os.IsNotExist(err) {
+			t.Error("corrupt legacy file should not have been deleted")
+		}
+	})
+
+	t.Run("new path takes precedence over legacy", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		legacyDir := filepath.Join(tmpDir, ".ollama", "config")
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(legacyDir, "config.json"), []byte(`{"integrations":{"old":{"models":["old-model"]}}}`), 0o644)
+
+		newDir := filepath.Join(tmpDir, ".ollama")
+		os.WriteFile(filepath.Join(newDir, "config.json"), []byte(`{"integrations":{"new":{"models":["new-model"]}}}`), 0o644)
+
+		cfg, err := load()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if _, ok := cfg.Integrations["new"]; !ok {
+			t.Error("expected new-path integration to be loaded")
+		}
+		if _, ok := cfg.Integrations["old"]; ok {
+			t.Error("legacy integration should not have been loaded")
+		}
+	})
+
+	t.Run("idempotent when called twice", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		legacyDir := filepath.Join(tmpDir, ".ollama", "config")
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(legacyDir, "config.json"), []byte(`{"integrations":{}}`), 0o644)
+
+		if _, err := migrateConfig(); err != nil {
+			t.Fatal(err)
+		}
+
+		migrated, err := migrateConfig()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if migrated {
+			t.Error("second migration should be a no-op")
+		}
+	})
+
+	t.Run("legacy directory preserved if not empty", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		legacyDir := filepath.Join(tmpDir, ".ollama", "config")
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(legacyDir, "config.json"), []byte(`{"integrations":{}}`), 0o644)
+		os.WriteFile(filepath.Join(legacyDir, "other-file.txt"), []byte("keep me"), 0o644)
+
+		if _, err := migrateConfig(); err != nil {
+			t.Fatal(err)
+		}
+
+		if _, err := os.Stat(legacyDir); os.IsNotExist(err) {
+			t.Error("directory with other files should not have been removed")
+		}
+		if _, err := os.Stat(filepath.Join(legacyDir, "other-file.txt")); os.IsNotExist(err) {
+			t.Error("other files in legacy directory should be untouched")
+		}
+	})
+
+	t.Run("save writes to new path after migration", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		legacyDir := filepath.Join(tmpDir, ".ollama", "config")
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(legacyDir, "config.json"), []byte(`{"integrations":{"claude":{"models":["llama3.2"]}}}`), 0o644)
+
+		// load triggers migration, then save should write to new path
+		if err := saveIntegration("codex", []string{"qwen2.5"}); err != nil {
+			t.Fatal(err)
+		}
+
+		newPath := filepath.Join(tmpDir, ".ollama", "config.json")
+		if _, err := os.Stat(newPath); os.IsNotExist(err) {
+			t.Error("save should write to new path")
+		}
+
+		// old path should not be recreated
+		if _, err := os.Stat(filepath.Join(legacyDir, "config.json")); !os.IsNotExist(err) {
+			t.Error("save should not recreate legacy path")
+		}
+	})
+
+	t.Run("load triggers migration transparently", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		legacyDir := filepath.Join(tmpDir, ".ollama", "config")
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(legacyDir, "config.json"), []byte(`{"integrations":{"claude":{"models":["llama3.2"]}}}`), 0o644)
+
+		cfg, err := load()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if cfg.Integrations["claude"] == nil || cfg.Integrations["claude"].Models[0] != "llama3.2" {
+			t.Error("migration via load() did not preserve data")
+		}
+	})
+}
+
+func TestSave(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("creates config file", func(t *testing.T) {
+		cfg := &config{
+			Integrations: map[string]*integration{
+				"test": {Models: []string{"model-a", "model-b"}},
+			},
+		}
+
+		if err := save(cfg); err != nil {
+			t.Fatal(err)
+		}
+
+		path, _ := configPath()
+		if _, err := os.Stat(path); os.IsNotExist(err) {
+			t.Error("config file was not created")
+		}
+	})
+
+	t.Run("round-trip preserves data", func(t *testing.T) {
+		cfg := &config{
+			Integrations: map[string]*integration{
+				"claude": {Models: []string{"llama3.2", "mistral"}},
+				"codex":  {Models: []string{"qwen2.5"}},
+			},
+		}
+
+		if err := save(cfg); err != nil {
+			t.Fatal(err)
+		}
+
+		loaded, err := load()
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if len(loaded.Integrations) != 2 {
+			t.Errorf("expected 2 integrations, got %d", len(loaded.Integrations))
+		}
+		if loaded.Integrations["claude"] == nil {
+			t.Error("missing claude integration")
+		}
+		if len(loaded.Integrations["claude"].Models) != 2 {
+			t.Errorf("expected 2 models for claude, got %d", len(loaded.Integrations["claude"].Models))
+		}
+	})
+}

cmd/config/droid.go

@@ -0,0 +1,196 @@
+package config
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"slices"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
+)
+
+// Droid implements Runner and Editor for Droid integration
+type Droid struct{}
+
+// droidSettings represents the Droid settings.json file (only fields we use)
+type droidSettings struct {
+	CustomModels           []modelEntry    `json:"customModels"`
+	SessionDefaultSettings sessionSettings `json:"sessionDefaultSettings"`
+}
+
+type sessionSettings struct {
+	Model           string `json:"model"`
+	ReasoningEffort string `json:"reasoningEffort"`
+}
+
+type modelEntry 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, args []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", args...)
+	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
+	}
+
+	// Read file once, unmarshal twice:
+	// map preserves unknown fields for writing back (including extra fields in model entries)
+	settingsMap := make(map[string]any)
+	var settings droidSettings
+	if data, err := os.ReadFile(settingsPath); err == nil {
+		if err := json.Unmarshal(data, &settingsMap); err != nil {
+			return fmt.Errorf("failed to parse settings file: %w, at: %s", err, settingsPath)
+		}
+		json.Unmarshal(data, &settings) // ignore error, zero values are fine
+	}
+
+	// Keep only non-Ollama models from the raw map (preserves extra fields)
+	// Rebuild Ollama models
+	var nonOllamaModels []any
+	if rawModels, ok := settingsMap["customModels"].([]any); ok {
+		for _, raw := range rawModels {
+			if m, ok := raw.(map[string]any); ok {
+				if m["apiKey"] != "ollama" {
+					nonOllamaModels = append(nonOllamaModels, raw)
+				}
+			}
+		}
+	}
+
+	// Build new Ollama model entries with sequential indices (0, 1, 2, ...)
+	client, _ := api.ClientFromEnvironment()
+
+	var newModels []any
+	var defaultModelID string
+	for i, model := range models {
+		maxOutput := 64000
+		if isCloudModel(context.Background(), client, model) {
+			if l, ok := lookupCloudModelLimit(model); ok {
+				maxOutput = l.Output
+			}
+		}
+		modelID := fmt.Sprintf("custom:%s-%d", model, i)
+		newModels = append(newModels, modelEntry{
+			Model:           model,
+			DisplayName:     model,
+			BaseURL:         envconfig.Host().String() + "/v1",
+			APIKey:          "ollama",
+			Provider:        "generic-chat-completion-api",
+			MaxOutputTokens: maxOutput,
+			SupportsImages:  false,
+			ID:              modelID,
+			Index:           i,
+		})
+		if i == 0 {
+			defaultModelID = modelID
+		}
+	}
+
+	settingsMap["customModels"] = append(newModels, nonOllamaModels...)
+
+	// Update session default settings (preserve unknown fields in the nested object)
+	sessionSettings, ok := settingsMap["sessionDefaultSettings"].(map[string]any)
+	if !ok {
+		sessionSettings = make(map[string]any)
+	}
+	sessionSettings["model"] = defaultModelID
+
+	if !isValidReasoningEffort(settings.SessionDefaultSettings.ReasoningEffort) {
+		sessionSettings["reasoningEffort"] = "none"
+	}
+
+	settingsMap["sessionDefaultSettings"] = sessionSettings
+
+	data, err := json.MarshalIndent(settingsMap, "", "  ")
+	if err != nil {
+		return err
+	}
+	return writeWithBackup(settingsPath, data)
+}
+
+func (d *Droid) Models() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+
+	data, err := os.ReadFile(filepath.Join(home, ".factory", "settings.json"))
+	if err != nil {
+		return nil
+	}
+
+	var settings droidSettings
+	if err := json.Unmarshal(data, &settings); err != nil {
+		return nil
+	}
+
+	var result []string
+	for _, m := range settings.CustomModels {
+		if m.APIKey == "ollama" {
+			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/files.go

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

cmd/config/files_test.go

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

cmd/config/openclaw.go

@@ -0,0 +1,252 @@
+package config
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+
+	"github.com/ollama/ollama/envconfig"
+)
+
+type Openclaw struct{}
+
+func (c *Openclaw) String() string { return "OpenClaw" }
+
+func (c *Openclaw) Run(model string, args []string) error {
+	bin := "openclaw"
+	if _, err := exec.LookPath(bin); err != nil {
+		bin = "clawdbot"
+		if _, err := exec.LookPath(bin); err != nil {
+			return fmt.Errorf("openclaw is not installed, install from https://docs.openclaw.ai")
+		}
+	}
+
+	models := []string{model}
+	if config, err := loadIntegration("openclaw"); err == nil && len(config.Models) > 0 {
+		models = config.Models
+	} else if config, err := loadIntegration("clawdbot"); err == nil && len(config.Models) > 0 {
+		models = config.Models
+	}
+	if err := c.Edit(models); err != nil {
+		return fmt.Errorf("setup failed: %w", err)
+	}
+
+	if !c.onboarded() {
+		// Onboarding not completed: run it (model already set via Edit)
+		// Use "ollama" as gateway token for simple local access
+		cmd := exec.Command(bin, "onboard",
+			"--auth-choice", "skip",
+			"--gateway-token", "ollama",
+		)
+		cmd.Stdin = os.Stdin
+		cmd.Stdout = os.Stdout
+		cmd.Stderr = os.Stderr
+		return cmd.Run()
+	}
+
+	// Onboarding completed: run gateway
+	cmd := exec.Command(bin, append([]string{"gateway"}, args...)...)
+	cmd.Stdin = os.Stdin
+
+	// Capture output to detect "already running" message
+	var outputBuf bytes.Buffer
+	cmd.Stdout = io.MultiWriter(os.Stdout, &outputBuf)
+	cmd.Stderr = io.MultiWriter(os.Stderr, &outputBuf)
+
+	err := cmd.Run()
+	if err != nil && strings.Contains(outputBuf.String(), "Gateway already running") {
+		fmt.Fprintf(os.Stderr, "%sOpenClaw has been configured with Ollama. Gateway is already running.%s\n", ansiGreen, ansiReset)
+		return nil
+	}
+	return err
+}
+
+// onboarded checks if OpenClaw onboarding wizard was completed
+// by looking for the wizard.lastRunAt marker in the config
+func (c *Openclaw) onboarded() bool {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return false
+	}
+
+	configPath := filepath.Join(home, ".openclaw", "openclaw.json")
+	legacyPath := filepath.Join(home, ".clawdbot", "clawdbot.json")
+
+	config := make(map[string]any)
+	if data, err := os.ReadFile(configPath); err == nil {
+		_ = json.Unmarshal(data, &config)
+	} else if data, err := os.ReadFile(legacyPath); err == nil {
+		_ = json.Unmarshal(data, &config)
+	} else {
+		return false
+	}
+
+	// Check for wizard.lastRunAt marker (set when onboarding completes)
+	wizard, _ := config["wizard"].(map[string]any)
+	if wizard == nil {
+		return false
+	}
+	lastRunAt, _ := wizard["lastRunAt"].(string)
+	return lastRunAt != ""
+}
+
+func (c *Openclaw) Paths() []string {
+	home, _ := os.UserHomeDir()
+	p := filepath.Join(home, ".openclaw", "openclaw.json")
+	if _, err := os.Stat(p); err == nil {
+		return []string{p}
+	}
+	legacy := filepath.Join(home, ".clawdbot", "clawdbot.json")
+	if _, err := os.Stat(legacy); err == nil {
+		return []string{legacy}
+	}
+	return nil
+}
+
+func (c *Openclaw) Edit(models []string) error {
+	if len(models) == 0 {
+		return nil
+	}
+
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return err
+	}
+
+	configPath := filepath.Join(home, ".openclaw", "openclaw.json")
+	legacyPath := filepath.Join(home, ".clawdbot", "clawdbot.json")
+	if err := os.MkdirAll(filepath.Dir(configPath), 0o755); err != nil {
+		return err
+	}
+
+	// Read into map[string]any to preserve unknown fields
+	config := make(map[string]any)
+	if data, err := os.ReadFile(configPath); err == nil {
+		_ = json.Unmarshal(data, &config)
+	} else if data, err := os.ReadFile(legacyPath); err == nil {
+		_ = json.Unmarshal(data, &config)
+	}
+
+	// Navigate/create: models.providers.ollama (preserving other providers)
+	modelsSection, _ := config["models"].(map[string]any)
+	if modelsSection == nil {
+		modelsSection = make(map[string]any)
+	}
+	providers, _ := modelsSection["providers"].(map[string]any)
+	if providers == nil {
+		providers = make(map[string]any)
+	}
+	ollama, _ := providers["ollama"].(map[string]any)
+	if ollama == nil {
+		ollama = make(map[string]any)
+	}
+
+	ollama["baseUrl"] = envconfig.Host().String() + "/v1"
+	// needed to register provider
+	ollama["apiKey"] = "ollama-local"
+	// TODO(parthsareen): potentially move to responses
+	ollama["api"] = "openai-completions"
+
+	// Build map of existing models to preserve user customizations
+	existingModels, _ := ollama["models"].([]any)
+	existingByID := make(map[string]map[string]any)
+	for _, m := range existingModels {
+		if entry, ok := m.(map[string]any); ok {
+			if id, ok := entry["id"].(string); ok {
+				existingByID[id] = entry
+			}
+		}
+	}
+
+	var newModels []any
+	for _, model := range models {
+		entry := map[string]any{
+			"id":        model,
+			"name":      model,
+			"reasoning": false,
+			"input":     []any{"text"},
+			"cost": map[string]any{
+				"input":      0,
+				"output":     0,
+				"cacheRead":  0,
+				"cacheWrite": 0,
+			},
+			// TODO(parthsareen): get these values from API
+			"contextWindow": 131072,
+			"maxTokens":     16384,
+		}
+		// Merge existing fields (user customizations)
+		if existing, ok := existingByID[model]; ok {
+			for k, v := range existing {
+				if _, isNew := entry[k]; !isNew {
+					entry[k] = v
+				}
+			}
+		}
+		newModels = append(newModels, entry)
+	}
+	ollama["models"] = newModels
+
+	providers["ollama"] = ollama
+	modelsSection["providers"] = providers
+	config["models"] = modelsSection
+
+	// Update agents.defaults.model.primary (preserving other agent settings)
+	agents, _ := config["agents"].(map[string]any)
+	if agents == nil {
+		agents = make(map[string]any)
+	}
+	defaults, _ := agents["defaults"].(map[string]any)
+	if defaults == nil {
+		defaults = make(map[string]any)
+	}
+	modelConfig, _ := defaults["model"].(map[string]any)
+	if modelConfig == nil {
+		modelConfig = make(map[string]any)
+	}
+	modelConfig["primary"] = "ollama/" + models[0]
+	defaults["model"] = modelConfig
+	agents["defaults"] = defaults
+	config["agents"] = agents
+
+	data, err := json.MarshalIndent(config, "", "  ")
+	if err != nil {
+		return err
+	}
+	return writeWithBackup(configPath, data)
+}
+
+func (c *Openclaw) Models() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+
+	config, err := readJSONFile(filepath.Join(home, ".openclaw", "openclaw.json"))
+	if err != nil {
+		config, err = readJSONFile(filepath.Join(home, ".clawdbot", "clawdbot.json"))
+		if err != nil {
+			return nil
+		}
+	}
+
+	modelsSection, _ := config["models"].(map[string]any)
+	providers, _ := modelsSection["providers"].(map[string]any)
+	ollama, _ := providers["ollama"].(map[string]any)
+	modelList, _ := ollama["models"].([]any)
+
+	var result []string
+	for _, m := range modelList {
+		if entry, ok := m.(map[string]any); ok {
+			if id, ok := entry["id"].(string); ok {
+				result = append(result, id)
+			}
+		}
+	}
+	return result
+}

cmd/config/openclaw_test.go

@@ -0,0 +1,878 @@
+package config
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestOpenclawIntegration(t *testing.T) {
+	c := &Openclaw{}
+
+	t.Run("String", func(t *testing.T) {
+		if got := c.String(); got != "OpenClaw" {
+			t.Errorf("String() = %q, want %q", got, "OpenClaw")
+		}
+	})
+
+	t.Run("implements Runner", func(t *testing.T) {
+		var _ Runner = c
+	})
+
+	t.Run("implements Editor", func(t *testing.T) {
+		var _ Editor = c
+	})
+}
+
+func TestOpenclawEdit(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".openclaw")
+	configPath := filepath.Join(configDir, "openclaw.json")
+
+	cleanup := func() { os.RemoveAll(configDir) }
+
+	t.Run("fresh install", func(t *testing.T) {
+		cleanup()
+		if err := c.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenclawModelExists(t, configPath, "llama3.2")
+		assertOpenclawPrimaryModel(t, configPath, "ollama/llama3.2")
+	})
+
+	t.Run("multiple models - first is primary", func(t *testing.T) {
+		cleanup()
+		if err := c.Edit([]string{"llama3.2", "mistral"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenclawModelExists(t, configPath, "llama3.2")
+		assertOpenclawModelExists(t, configPath, "mistral")
+		assertOpenclawPrimaryModel(t, configPath, "ollama/llama3.2")
+	})
+
+	t.Run("preserve other providers", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"models":{"providers":{"anthropic":{"apiKey":"xxx"}}}}`), 0o644)
+		if err := c.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		models := cfg["models"].(map[string]any)
+		providers := models["providers"].(map[string]any)
+		if providers["anthropic"] == nil {
+			t.Error("anthropic provider was removed")
+		}
+	})
+
+	t.Run("preserve top-level keys", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"theme":"dark","mcp":{"servers":{}}}`), 0o644)
+		if err := c.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["mcp"] == nil {
+			t.Error("mcp was removed")
+		}
+	})
+
+	t.Run("preserve user customizations on models", func(t *testing.T) {
+		cleanup()
+		c.Edit([]string{"llama3.2"})
+
+		// User adds custom field
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		models := cfg["models"].(map[string]any)
+		providers := models["providers"].(map[string]any)
+		ollama := providers["ollama"].(map[string]any)
+		modelList := ollama["models"].([]any)
+		entry := modelList[0].(map[string]any)
+		entry["customField"] = "user-value"
+		configData, _ := json.MarshalIndent(cfg, "", "  ")
+		os.WriteFile(configPath, configData, 0o644)
+
+		// Re-run Edit
+		c.Edit([]string{"llama3.2"})
+
+		data, _ = os.ReadFile(configPath)
+		json.Unmarshal(data, &cfg)
+		models = cfg["models"].(map[string]any)
+		providers = models["providers"].(map[string]any)
+		ollama = providers["ollama"].(map[string]any)
+		modelList = ollama["models"].([]any)
+		entry = modelList[0].(map[string]any)
+		if entry["customField"] != "user-value" {
+			t.Error("custom field was lost")
+		}
+	})
+
+	t.Run("edit replaces models list", func(t *testing.T) {
+		cleanup()
+		c.Edit([]string{"llama3.2", "mistral"})
+		c.Edit([]string{"llama3.2"})
+
+		assertOpenclawModelExists(t, configPath, "llama3.2")
+		assertOpenclawModelNotExists(t, configPath, "mistral")
+	})
+
+	t.Run("empty models is no-op", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		original := `{"existing":"data"}`
+		os.WriteFile(configPath, []byte(original), 0o644)
+
+		c.Edit([]string{})
+
+		data, _ := os.ReadFile(configPath)
+		if string(data) != original {
+			t.Error("empty models should not modify file")
+		}
+	})
+
+	t.Run("corrupted JSON treated as empty", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{corrupted`), 0o644)
+
+		if err := c.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		if err := json.Unmarshal(data, &cfg); err != nil {
+			t.Error("result should be valid JSON")
+		}
+	})
+
+	t.Run("wrong type models section", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"models":"not a map"}`), 0o644)
+
+		if err := c.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenclawModelExists(t, configPath, "llama3.2")
+	})
+}
+
+func TestOpenclawModels(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	t.Run("no config returns nil", func(t *testing.T) {
+		if models := c.Models(); len(models) > 0 {
+			t.Errorf("expected nil/empty, got %v", models)
+		}
+	})
+
+	t.Run("returns all ollama models", func(t *testing.T) {
+		configDir := filepath.Join(tmpDir, ".openclaw")
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(filepath.Join(configDir, "openclaw.json"), []byte(`{
+			"models":{"providers":{"ollama":{"models":[
+				{"id":"llama3.2"},
+				{"id":"mistral"}
+			]}}}
+		}`), 0o644)
+
+		models := c.Models()
+		if len(models) != 2 {
+			t.Errorf("expected 2 models, got %v", models)
+		}
+	})
+}
+
+// Helper functions
+func assertOpenclawModelExists(t *testing.T, path, model string) {
+	t.Helper()
+	data, _ := os.ReadFile(path)
+	var cfg map[string]any
+	json.Unmarshal(data, &cfg)
+	models := cfg["models"].(map[string]any)
+	providers := models["providers"].(map[string]any)
+	ollama := providers["ollama"].(map[string]any)
+	modelList := ollama["models"].([]any)
+	for _, m := range modelList {
+		if entry, ok := m.(map[string]any); ok {
+			if entry["id"] == model {
+				return
+			}
+		}
+	}
+	t.Errorf("model %s not found", model)
+}
+
+func assertOpenclawModelNotExists(t *testing.T, path, model string) {
+	t.Helper()
+	data, _ := os.ReadFile(path)
+	var cfg map[string]any
+	json.Unmarshal(data, &cfg)
+	models, _ := cfg["models"].(map[string]any)
+	providers, _ := models["providers"].(map[string]any)
+	ollama, _ := providers["ollama"].(map[string]any)
+	modelList, _ := ollama["models"].([]any)
+	for _, m := range modelList {
+		if entry, ok := m.(map[string]any); ok {
+			if entry["id"] == model {
+				t.Errorf("model %s should not exist", model)
+			}
+		}
+	}
+}
+
+func assertOpenclawPrimaryModel(t *testing.T, path, expected string) {
+	t.Helper()
+	data, _ := os.ReadFile(path)
+	var cfg map[string]any
+	json.Unmarshal(data, &cfg)
+	agents := cfg["agents"].(map[string]any)
+	defaults := agents["defaults"].(map[string]any)
+	model := defaults["model"].(map[string]any)
+	if model["primary"] != expected {
+		t.Errorf("primary model = %v, want %v", model["primary"], expected)
+	}
+}
+
+func TestOpenclawPaths(t *testing.T) {
+	c := &Openclaw{}
+
+	t.Run("returns path when config exists", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		configDir := filepath.Join(tmpDir, ".openclaw")
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(filepath.Join(configDir, "openclaw.json"), []byte(`{}`), 0o644)
+
+		paths := c.Paths()
+		if len(paths) != 1 {
+			t.Errorf("expected 1 path, got %d", len(paths))
+		}
+	})
+
+	t.Run("returns nil when config missing", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		if paths := c.Paths(); paths != nil {
+			t.Errorf("expected nil, got %v", paths)
+		}
+	})
+}
+
+func TestOpenclawModelsEdgeCases(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+	configDir := filepath.Join(tmpDir, ".openclaw")
+	configPath := filepath.Join(configDir, "openclaw.json")
+	cleanup := func() { os.RemoveAll(configDir) }
+
+	t.Run("corrupted JSON returns nil", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{corrupted`), 0o644)
+		if models := c.Models(); models != nil {
+			t.Errorf("expected nil, got %v", models)
+		}
+	})
+
+	t.Run("wrong type at models level", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"models":"string"}`), 0o644)
+		if models := c.Models(); models != nil {
+			t.Errorf("expected nil, got %v", models)
+		}
+	})
+
+	t.Run("wrong type at providers level", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"models":{"providers":"string"}}`), 0o644)
+		if models := c.Models(); models != nil {
+			t.Errorf("expected nil, got %v", models)
+		}
+	})
+
+	t.Run("wrong type at ollama level", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"models":{"providers":{"ollama":"string"}}}`), 0o644)
+		if models := c.Models(); models != nil {
+			t.Errorf("expected nil, got %v", models)
+		}
+	})
+
+	t.Run("model entry missing id", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"models":{"providers":{"ollama":{"models":[{"name":"test"}]}}}}`), 0o644)
+		if len(c.Models()) != 0 {
+			t.Error("expected empty for missing id")
+		}
+	})
+
+	t.Run("model id is not string", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"models":{"providers":{"ollama":{"models":[{"id":123}]}}}}`), 0o644)
+		if len(c.Models()) != 0 {
+			t.Error("expected empty for non-string id")
+		}
+	})
+}
+
+func TestOpenclawEditSchemaFields(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+	configPath := filepath.Join(tmpDir, ".openclaw", "openclaw.json")
+
+	if err := c.Edit([]string{"llama3.2"}); err != nil {
+		t.Fatal(err)
+	}
+
+	data, _ := os.ReadFile(configPath)
+	var cfg map[string]any
+	json.Unmarshal(data, &cfg)
+	models := cfg["models"].(map[string]any)
+	providers := models["providers"].(map[string]any)
+	ollama := providers["ollama"].(map[string]any)
+	modelList := ollama["models"].([]any)
+	entry := modelList[0].(map[string]any)
+
+	// Verify required schema fields
+	if entry["reasoning"] != false {
+		t.Error("reasoning should be false")
+	}
+	if entry["input"] == nil {
+		t.Error("input should be set")
+	}
+	if entry["contextWindow"] == nil {
+		t.Error("contextWindow should be set")
+	}
+	if entry["maxTokens"] == nil {
+		t.Error("maxTokens should be set")
+	}
+	cost := entry["cost"].(map[string]any)
+	if cost["cacheRead"] == nil {
+		t.Error("cost.cacheRead should be set")
+	}
+	if cost["cacheWrite"] == nil {
+		t.Error("cost.cacheWrite should be set")
+	}
+}
+
+func TestOpenclawEditModelNames(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+	configPath := filepath.Join(tmpDir, ".openclaw", "openclaw.json")
+	cleanup := func() { os.RemoveAll(filepath.Join(tmpDir, ".openclaw")) }
+
+	t.Run("model with colon tag", func(t *testing.T) {
+		cleanup()
+		if err := c.Edit([]string{"llama3.2:70b"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenclawModelExists(t, configPath, "llama3.2:70b")
+		assertOpenclawPrimaryModel(t, configPath, "ollama/llama3.2:70b")
+	})
+
+	t.Run("model with slash", func(t *testing.T) {
+		cleanup()
+		if err := c.Edit([]string{"library/model:tag"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenclawModelExists(t, configPath, "library/model:tag")
+		assertOpenclawPrimaryModel(t, configPath, "ollama/library/model:tag")
+	})
+
+	t.Run("model with hyphen", func(t *testing.T) {
+		cleanup()
+		if err := c.Edit([]string{"test-model"}); err != nil {
+			t.Fatal(err)
+		}
+		assertOpenclawModelExists(t, configPath, "test-model")
+	})
+}
+
+func TestOpenclawEditAgentsPreservation(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+	configDir := filepath.Join(tmpDir, ".openclaw")
+	configPath := filepath.Join(configDir, "openclaw.json")
+	cleanup := func() { os.RemoveAll(configDir) }
+
+	t.Run("preserve other agent defaults", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"agents":{"defaults":{"model":{"primary":"old"},"temperature":0.7}}}`), 0o644)
+
+		c.Edit([]string{"llama3.2"})
+
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		agents := cfg["agents"].(map[string]any)
+		defaults := agents["defaults"].(map[string]any)
+		if defaults["temperature"] != 0.7 {
+			t.Error("temperature setting was lost")
+		}
+	})
+
+	t.Run("preserve other agents besides defaults", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"agents":{"defaults":{},"custom-agent":{"foo":"bar"}}}`), 0o644)
+
+		c.Edit([]string{"llama3.2"})
+
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		agents := cfg["agents"].(map[string]any)
+		if agents["custom-agent"] == nil {
+			t.Error("custom-agent was lost")
+		}
+	})
+}
+
+const testOpenclawFixture = `{
+  "theme": "dark",
+  "mcp": {"servers": {"custom": {"enabled": true}}},
+  "models": {
+    "providers": {
+      "anthropic": {"apiKey": "xxx"},
+      "ollama": {
+        "baseUrl": "http://127.0.0.1:11434/v1",
+        "models": [{"id": "old-model", "customField": "preserved"}]
+      }
+    }
+  },
+  "agents": {
+    "defaults": {"model": {"primary": "old"}, "temperature": 0.7},
+    "custom-agent": {"foo": "bar"}
+  }
+}`
+
+func TestOpenclawEdit_RoundTrip(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+	configDir := filepath.Join(tmpDir, ".openclaw")
+	configPath := filepath.Join(configDir, "openclaw.json")
+
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(testOpenclawFixture), 0o644)
+
+	if err := c.Edit([]string{"llama3.2", "mistral"}); err != nil {
+		t.Fatal(err)
+	}
+
+	data, _ := os.ReadFile(configPath)
+	var cfg map[string]any
+	json.Unmarshal(data, &cfg)
+
+	// Verify top-level preserved
+	if cfg["theme"] != "dark" {
+		t.Error("theme not preserved")
+	}
+	mcp := cfg["mcp"].(map[string]any)
+	servers := mcp["servers"].(map[string]any)
+	if servers["custom"] == nil {
+		t.Error("mcp.servers.custom not preserved")
+	}
+
+	// Verify other providers preserved
+	models := cfg["models"].(map[string]any)
+	providers := models["providers"].(map[string]any)
+	if providers["anthropic"] == nil {
+		t.Error("anthropic provider not preserved")
+	}
+
+	// Verify agents preserved
+	agents := cfg["agents"].(map[string]any)
+	if agents["custom-agent"] == nil {
+		t.Error("custom-agent not preserved")
+	}
+	defaults := agents["defaults"].(map[string]any)
+	if defaults["temperature"] != 0.7 {
+		t.Error("temperature not preserved")
+	}
+}
+
+func TestOpenclawEdit_Idempotent(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+	configDir := filepath.Join(tmpDir, ".openclaw")
+	configPath := filepath.Join(configDir, "openclaw.json")
+
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(testOpenclawFixture), 0o644)
+
+	c.Edit([]string{"llama3.2", "mistral"})
+	firstData, _ := os.ReadFile(configPath)
+
+	c.Edit([]string{"llama3.2", "mistral"})
+	secondData, _ := os.ReadFile(configPath)
+
+	if string(firstData) != string(secondData) {
+		t.Error("repeated edits with same models produced different results")
+	}
+}
+
+func TestOpenclawEdit_MultipleConsecutiveEdits(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+	configDir := filepath.Join(tmpDir, ".openclaw")
+	configPath := filepath.Join(configDir, "openclaw.json")
+
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(testOpenclawFixture), 0o644)
+
+	for i := range 10 {
+		models := []string{"model-a", "model-b"}
+		if i%2 == 0 {
+			models = []string{"model-x", "model-y", "model-z"}
+		}
+		if err := c.Edit(models); err != nil {
+			t.Fatalf("edit %d failed: %v", i, err)
+		}
+	}
+
+	data, _ := os.ReadFile(configPath)
+	var cfg map[string]any
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Fatalf("file is not valid JSON after multiple edits: %v", err)
+	}
+
+	if cfg["theme"] != "dark" {
+		t.Error("theme lost after multiple edits")
+	}
+}
+
+func TestOpenclawEdit_BackupCreated(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+	configDir := filepath.Join(tmpDir, ".openclaw")
+	configPath := filepath.Join(configDir, "openclaw.json")
+	backupDir := filepath.Join(os.TempDir(), "ollama-backups")
+
+	os.MkdirAll(configDir, 0o755)
+	uniqueMarker := fmt.Sprintf("test-marker-%d", os.Getpid())
+	original := fmt.Sprintf(`{"theme": "%s"}`, uniqueMarker)
+	os.WriteFile(configPath, []byte(original), 0o644)
+
+	if err := c.Edit([]string{"model-a"}); err != nil {
+		t.Fatal(err)
+	}
+
+	backups, _ := filepath.Glob(filepath.Join(backupDir, "openclaw.json.*"))
+	foundBackup := false
+	for _, backup := range backups {
+		data, _ := os.ReadFile(backup)
+		if string(data) == original {
+			foundBackup = true
+			break
+		}
+	}
+
+	if !foundBackup {
+		t.Error("backup with original content not found")
+	}
+}
+
+func TestOpenclawClawdbotAlias(t *testing.T) {
+	for _, alias := range []string{"clawdbot", "moltbot"} {
+		t.Run(alias+" alias resolves to Openclaw runner", func(t *testing.T) {
+			r, ok := integrations[alias]
+			if !ok {
+				t.Fatalf("%s not found in integrations", alias)
+			}
+			if _, ok := r.(*Openclaw); !ok {
+				t.Errorf("%s integration is %T, want *Openclaw", alias, r)
+			}
+		})
+
+		t.Run(alias+" is hidden from selector", func(t *testing.T) {
+			if !integrationAliases[alias] {
+				t.Errorf("%s should be in integrationAliases", alias)
+			}
+		})
+	}
+}
+
+func TestOpenclawLegacyPaths(t *testing.T) {
+	c := &Openclaw{}
+
+	t.Run("falls back to legacy clawdbot path", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		legacyDir := filepath.Join(tmpDir, ".clawdbot")
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(legacyDir, "clawdbot.json"), []byte(`{}`), 0o644)
+
+		paths := c.Paths()
+		if len(paths) != 1 {
+			t.Fatalf("expected 1 path, got %d", len(paths))
+		}
+		if paths[0] != filepath.Join(legacyDir, "clawdbot.json") {
+			t.Errorf("expected legacy path, got %s", paths[0])
+		}
+	})
+
+	t.Run("prefers new path over legacy", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		newDir := filepath.Join(tmpDir, ".openclaw")
+		legacyDir := filepath.Join(tmpDir, ".clawdbot")
+		os.MkdirAll(newDir, 0o755)
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(newDir, "openclaw.json"), []byte(`{}`), 0o644)
+		os.WriteFile(filepath.Join(legacyDir, "clawdbot.json"), []byte(`{}`), 0o644)
+
+		paths := c.Paths()
+		if len(paths) != 1 {
+			t.Fatalf("expected 1 path, got %d", len(paths))
+		}
+		if paths[0] != filepath.Join(newDir, "openclaw.json") {
+			t.Errorf("expected new path, got %s", paths[0])
+		}
+	})
+
+	t.Run("Models reads from legacy path", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		legacyDir := filepath.Join(tmpDir, ".clawdbot")
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(legacyDir, "clawdbot.json"), []byte(`{
+			"models":{"providers":{"ollama":{"models":[{"id":"llama3.2"}]}}}
+		}`), 0o644)
+
+		models := c.Models()
+		if len(models) != 1 || models[0] != "llama3.2" {
+			t.Errorf("expected [llama3.2], got %v", models)
+		}
+	})
+
+	t.Run("Models prefers new path over legacy", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		newDir := filepath.Join(tmpDir, ".openclaw")
+		legacyDir := filepath.Join(tmpDir, ".clawdbot")
+		os.MkdirAll(newDir, 0o755)
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(newDir, "openclaw.json"), []byte(`{
+			"models":{"providers":{"ollama":{"models":[{"id":"new-model"}]}}}
+		}`), 0o644)
+		os.WriteFile(filepath.Join(legacyDir, "clawdbot.json"), []byte(`{
+			"models":{"providers":{"ollama":{"models":[{"id":"legacy-model"}]}}}
+		}`), 0o644)
+
+		models := c.Models()
+		if len(models) != 1 || models[0] != "new-model" {
+			t.Errorf("expected [new-model], got %v", models)
+		}
+	})
+
+	t.Run("Edit reads new path over legacy when both exist", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		newDir := filepath.Join(tmpDir, ".openclaw")
+		legacyDir := filepath.Join(tmpDir, ".clawdbot")
+		os.MkdirAll(newDir, 0o755)
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(newDir, "openclaw.json"), []byte(`{"theme":"new"}`), 0o644)
+		os.WriteFile(filepath.Join(legacyDir, "clawdbot.json"), []byte(`{"theme":"legacy"}`), 0o644)
+
+		if err := c.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+
+		data, _ := os.ReadFile(filepath.Join(newDir, "openclaw.json"))
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		if cfg["theme"] != "new" {
+			t.Errorf("expected theme from new config, got %v", cfg["theme"])
+		}
+	})
+
+	t.Run("Edit migrates from legacy config", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		legacyDir := filepath.Join(tmpDir, ".clawdbot")
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(legacyDir, "clawdbot.json"), []byte(`{"theme":"dark"}`), 0o644)
+
+		if err := c.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Should write to new path
+		newPath := filepath.Join(tmpDir, ".openclaw", "openclaw.json")
+		data, err := os.ReadFile(newPath)
+		if err != nil {
+			t.Fatal("expected new config file to be created")
+		}
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		if cfg["theme"] != "dark" {
+			t.Error("legacy theme setting was not migrated")
+		}
+	})
+}
+
+func TestOpenclawEdit_CreatesDirectoryIfMissing(t *testing.T) {
+	c := &Openclaw{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+	configDir := filepath.Join(tmpDir, ".openclaw")
+
+	if _, err := os.Stat(configDir); !os.IsNotExist(err) {
+		t.Fatal("directory should not exist before test")
+	}
+
+	if err := c.Edit([]string{"model-a"}); err != nil {
+		t.Fatal(err)
+	}
+
+	if _, err := os.Stat(configDir); os.IsNotExist(err) {
+		t.Fatal("directory was not created")
+	}
+}
+
+func TestOpenclawOnboarded(t *testing.T) {
+	c := &Openclaw{}
+
+	t.Run("returns false when no config exists", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		if c.onboarded() {
+			t.Error("expected false when no config exists")
+		}
+	})
+
+	t.Run("returns false when config exists but no wizard section", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		configDir := filepath.Join(tmpDir, ".openclaw")
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(filepath.Join(configDir, "openclaw.json"), []byte(`{"theme":"dark"}`), 0o644)
+
+		if c.onboarded() {
+			t.Error("expected false when no wizard section")
+		}
+	})
+
+	t.Run("returns false when wizard section exists but no lastRunAt", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		configDir := filepath.Join(tmpDir, ".openclaw")
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(filepath.Join(configDir, "openclaw.json"), []byte(`{"wizard":{}}`), 0o644)
+
+		if c.onboarded() {
+			t.Error("expected false when wizard.lastRunAt is missing")
+		}
+	})
+
+	t.Run("returns false when wizard.lastRunAt is empty string", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		configDir := filepath.Join(tmpDir, ".openclaw")
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(filepath.Join(configDir, "openclaw.json"), []byte(`{"wizard":{"lastRunAt":""}}`), 0o644)
+
+		if c.onboarded() {
+			t.Error("expected false when wizard.lastRunAt is empty")
+		}
+	})
+
+	t.Run("returns true when wizard.lastRunAt is set", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		configDir := filepath.Join(tmpDir, ".openclaw")
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(filepath.Join(configDir, "openclaw.json"), []byte(`{"wizard":{"lastRunAt":"2024-01-01T00:00:00Z"}}`), 0o644)
+
+		if !c.onboarded() {
+			t.Error("expected true when wizard.lastRunAt is set")
+		}
+	})
+
+	t.Run("checks legacy clawdbot path", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		legacyDir := filepath.Join(tmpDir, ".clawdbot")
+		os.MkdirAll(legacyDir, 0o755)
+		os.WriteFile(filepath.Join(legacyDir, "clawdbot.json"), []byte(`{"wizard":{"lastRunAt":"2024-01-01T00:00:00Z"}}`), 0o644)
+
+		if !c.onboarded() {
+			t.Error("expected true when legacy config has wizard.lastRunAt")
+		}
+	})
+
+	t.Run("prefers new path over legacy", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		newDir := filepath.Join(tmpDir, ".openclaw")
+		legacyDir := filepath.Join(tmpDir, ".clawdbot")
+		os.MkdirAll(newDir, 0o755)
+		os.MkdirAll(legacyDir, 0o755)
+		// New path has no wizard marker
+		os.WriteFile(filepath.Join(newDir, "openclaw.json"), []byte(`{}`), 0o644)
+		// Legacy has wizard marker
+		os.WriteFile(filepath.Join(legacyDir, "clawdbot.json"), []byte(`{"wizard":{"lastRunAt":"2024-01-01T00:00:00Z"}}`), 0o644)
+
+		if c.onboarded() {
+			t.Error("expected false - should prefer new path which has no wizard marker")
+		}
+	})
+
+	t.Run("handles corrupted JSON gracefully", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		configDir := filepath.Join(tmpDir, ".openclaw")
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(filepath.Join(configDir, "openclaw.json"), []byte(`{corrupted`), 0o644)
+
+		if c.onboarded() {
+			t.Error("expected false for corrupted JSON")
+		}
+	})
+
+	t.Run("handles wrong type for wizard section", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+		configDir := filepath.Join(tmpDir, ".openclaw")
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(filepath.Join(configDir, "openclaw.json"), []byte(`{"wizard":"not a map"}`), 0o644)
+
+		if c.onboarded() {
+			t.Error("expected false when wizard is wrong type")
+		}
+	})
+}

cmd/config/opencode.go

@@ -0,0 +1,287 @@
+package config
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"maps"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
+)
+
+// OpenCode implements Runner and Editor for OpenCode integration
+type OpenCode struct{}
+
+// cloudModelLimit holds context and output token limits for a cloud model.
+type cloudModelLimit struct {
+	Context int
+	Output  int
+}
+
+// cloudModelLimits maps cloud model base names to their token limits.
+// TODO(parthsareen): grab context/output limits from model info instead of hardcoding
+var cloudModelLimits = map[string]cloudModelLimit{
+	"cogito-2.1:671b":     {Context: 163_840, Output: 65_536},
+	"deepseek-v3.1:671b":  {Context: 163_840, Output: 163_840},
+	"deepseek-v3.2":       {Context: 163_840, Output: 65_536},
+	"glm-4.6":             {Context: 202_752, Output: 131_072},
+	"glm-4.7":             {Context: 202_752, Output: 131_072},
+	"gpt-oss:120b":        {Context: 131_072, Output: 131_072},
+	"gpt-oss:20b":         {Context: 131_072, Output: 131_072},
+	"kimi-k2:1t":          {Context: 262_144, Output: 262_144},
+	"kimi-k2.5":           {Context: 262_144, Output: 262_144},
+	"kimi-k2-thinking":    {Context: 262_144, Output: 262_144},
+	"nemotron-3-nano:30b": {Context: 1_048_576, Output: 131_072},
+	"qwen3-coder:480b":    {Context: 262_144, Output: 65_536},
+	"qwen3-coder-next":    {Context: 262_144, Output: 32_768},
+	"qwen3-next:80b":      {Context: 262_144, Output: 32_768},
+}
+
+// lookupCloudModelLimit returns the token limits for a cloud model.
+// It tries the exact name first, then strips the ":cloud" suffix.
+func lookupCloudModelLimit(name string) (cloudModelLimit, bool) {
+	if l, ok := cloudModelLimits[name]; ok {
+		return l, true
+	}
+	base := strings.TrimSuffix(name, ":cloud")
+	if base != name {
+		if l, ok := cloudModelLimits[base]; ok {
+			return l, true
+		}
+	}
+	return cloudModelLimit{}, false
+}
+
+func (o *OpenCode) String() string { return "OpenCode" }
+
+func (o *OpenCode) Run(model string, args []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", args...)
+	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": envconfig.Host().String() + "/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 isOllamaModel(cfgMap) && !selectedSet[name] {
+				delete(models, name)
+			}
+		}
+	}
+
+	client, _ := api.ClientFromEnvironment()
+
+	for _, model := range modelList {
+		if existing, ok := models[model].(map[string]any); ok {
+			// migrate existing models without _launch marker
+			if isOllamaModel(existing) {
+				existing["_launch"] = true
+				if name, ok := existing["name"].(string); ok {
+					existing["name"] = strings.TrimSuffix(name, " [Ollama]")
+				}
+			}
+			if isCloudModel(context.Background(), client, model) {
+				if l, ok := lookupCloudModelLimit(model); ok {
+					existing["limit"] = map[string]any{
+						"context": l.Context,
+						"output":  l.Output,
+					}
+				}
+			}
+			continue
+		}
+		entry := map[string]any{
+			"name":    model,
+			"_launch": true,
+		}
+		if isCloudModel(context.Background(), client, model) {
+			if l, ok := lookupCloudModelLimit(model); ok {
+				entry["limit"] = map[string]any{
+					"context": l.Context,
+					"output":  l.Output,
+				}
+			}
+		}
+		models[model] = entry
+	}
+
+	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
+}
+
+// isOllamaModel reports whether a model config entry is managed by us
+func isOllamaModel(cfg map[string]any) bool {
+	if v, ok := cfg["_launch"].(bool); ok && v {
+		return true
+	}
+	// previously used [Ollama] as a suffix for the model managed by ollama launch
+	if name, ok := cfg["name"].(string); ok {
+		return strings.HasSuffix(name, "[Ollama]")
+	}
+	return false
+}

cmd/config/opencode_test.go

@@ -0,0 +1,668 @@
+package config
+
+import (
+	"encoding/json"
+	"fmt"
+	"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("preserve user customizations on managed models", func(t *testing.T) {
+		cleanup()
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Add custom fields to the model entry (simulating user edits)
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		provider := cfg["provider"].(map[string]any)
+		ollama := provider["ollama"].(map[string]any)
+		models := ollama["models"].(map[string]any)
+		entry := models["llama3.2"].(map[string]any)
+		entry["_myPref"] = "custom-value"
+		entry["_myNum"] = 42
+		configData, _ := json.MarshalIndent(cfg, "", "  ")
+		os.WriteFile(configPath, configData, 0o644)
+
+		// Re-run Edit — should preserve custom fields
+		if err := o.Edit([]string{"llama3.2"}); err != nil {
+			t.Fatal(err)
+		}
+
+		data, _ = os.ReadFile(configPath)
+		json.Unmarshal(data, &cfg)
+		provider = cfg["provider"].(map[string]any)
+		ollama = provider["ollama"].(map[string]any)
+		models = ollama["models"].(map[string]any)
+		entry = models["llama3.2"].(map[string]any)
+
+		if entry["_myPref"] != "custom-value" {
+			t.Errorf("_myPref was lost: got %v", entry["_myPref"])
+		}
+		if entry["_myNum"] != float64(42) {
+			t.Errorf("_myNum was lost: got %v", entry["_myNum"])
+		}
+		if v, ok := entry["_launch"].(bool); !ok || !v {
+			t.Errorf("_launch marker missing or false: got %v", entry["_launch"])
+		}
+	})
+
+	t.Run("migrate legacy [Ollama] suffix entries", func(t *testing.T) {
+		cleanup()
+		// Write a config with a legacy entry (has [Ollama] suffix but no _launch marker)
+		os.MkdirAll(configDir, 0o755)
+		os.WriteFile(configPath, []byte(`{"provider":{"ollama":{"models":{"llama3.2":{"name":"llama3.2 [Ollama]"}}}}}`), 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)
+		ollama := provider["ollama"].(map[string]any)
+		models := ollama["models"].(map[string]any)
+		entry := models["llama3.2"].(map[string]any)
+
+		// _launch marker should be added
+		if v, ok := entry["_launch"].(bool); !ok || !v {
+			t.Errorf("_launch marker not added during migration: got %v", entry["_launch"])
+		}
+		// [Ollama] suffix should be stripped
+		if name, ok := entry["name"].(string); !ok || name != "llama3.2" {
+			t.Errorf("name suffix not stripped: got %q", entry["name"])
+		}
+	})
+
+	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 readOpenCodeModel(t *testing.T, configPath, model string) map[string]any {
+	t.Helper()
+	data, err := os.ReadFile(configPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var cfg map[string]any
+	json.Unmarshal(data, &cfg)
+	provider := cfg["provider"].(map[string]any)
+	ollama := provider["ollama"].(map[string]any)
+	models := ollama["models"].(map[string]any)
+	entry, ok := models[model].(map[string]any)
+	if !ok {
+		t.Fatalf("model %s not found in config", model)
+	}
+	return entry
+}
+
+func TestOpenCodeEdit_LocalModelNoLimit(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configPath := filepath.Join(tmpDir, ".config", "opencode", "opencode.json")
+
+	if err := o.Edit([]string{"llama3.2"}); err != nil {
+		t.Fatal(err)
+	}
+
+	entry := readOpenCodeModel(t, configPath, "llama3.2")
+	if entry["limit"] != nil {
+		t.Errorf("local model should not have limit set, got %v", entry["limit"])
+	}
+}
+
+func TestOpenCodeEdit_PreservesUserLimit(t *testing.T) {
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+
+	// Set up a model with a user-configured limit
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(`{
+		"provider": {
+			"ollama": {
+				"models": {
+					"llama3.2": {
+						"name": "llama3.2",
+						"_launch": true,
+						"limit": {"context": 8192, "output": 4096}
+					}
+				}
+			}
+		}
+	}`), 0o644)
+
+	// Re-edit should preserve the user's limit (not delete it)
+	if err := o.Edit([]string{"llama3.2"}); err != nil {
+		t.Fatal(err)
+	}
+
+	entry := readOpenCodeModel(t, configPath, "llama3.2")
+	limit, ok := entry["limit"].(map[string]any)
+	if !ok {
+		t.Fatal("user-configured limit was removed")
+	}
+	if limit["context"] != float64(8192) {
+		t.Errorf("context limit changed: got %v, want 8192", limit["context"])
+	}
+	if limit["output"] != float64(4096) {
+		t.Errorf("output limit changed: got %v, want 4096", limit["output"])
+	}
+}
+
+func TestOpenCodeEdit_CloudModelLimitStructure(t *testing.T) {
+	// Verify that when a cloud model entry has limits set (as Edit would do),
+	// the structure matches what opencode expects and re-edit preserves them.
+	o := &OpenCode{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".config", "opencode")
+	configPath := filepath.Join(configDir, "opencode.json")
+
+	expected := cloudModelLimits["glm-4.7"]
+
+	// Simulate a cloud model that already has the limit set by a previous Edit
+	os.MkdirAll(configDir, 0o755)
+	os.WriteFile(configPath, []byte(fmt.Sprintf(`{
+		"provider": {
+			"ollama": {
+				"models": {
+					"glm-4.7:cloud": {
+						"name": "glm-4.7:cloud",
+						"_launch": true,
+						"limit": {"context": %d, "output": %d}
+					}
+				}
+			}
+		}
+	}`, expected.Context, expected.Output)), 0o644)
+
+	// Re-edit should preserve the cloud model limit
+	if err := o.Edit([]string{"glm-4.7:cloud"}); err != nil {
+		t.Fatal(err)
+	}
+
+	entry := readOpenCodeModel(t, configPath, "glm-4.7:cloud")
+	limit, ok := entry["limit"].(map[string]any)
+	if !ok {
+		t.Fatal("cloud model limit was removed on re-edit")
+	}
+	if limit["context"] != float64(expected.Context) {
+		t.Errorf("context = %v, want %d", limit["context"], expected.Context)
+	}
+	if limit["output"] != float64(expected.Output) {
+		t.Errorf("output = %v, want %d", limit["output"], expected.Output)
+	}
+}
+
+func TestLookupCloudModelLimit(t *testing.T) {
+	tests := []struct {
+		name        string
+		wantOK      bool
+		wantContext int
+		wantOutput  int
+	}{
+		{"glm-4.7", true, 202_752, 131_072},
+		{"glm-4.7:cloud", true, 202_752, 131_072},
+		{"kimi-k2.5", true, 262_144, 262_144},
+		{"kimi-k2.5:cloud", true, 262_144, 262_144},
+		{"deepseek-v3.2", true, 163_840, 65_536},
+		{"deepseek-v3.2:cloud", true, 163_840, 65_536},
+		{"qwen3-coder:480b", true, 262_144, 65_536},
+		{"qwen3-coder-next:cloud", true, 262_144, 32_768},
+		{"llama3.2", false, 0, 0},
+		{"unknown-model:cloud", false, 0, 0},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			l, ok := lookupCloudModelLimit(tt.name)
+			if ok != tt.wantOK {
+				t.Errorf("lookupCloudModelLimit(%q) ok = %v, want %v", tt.name, ok, tt.wantOK)
+			}
+			if ok {
+				if l.Context != tt.wantContext {
+					t.Errorf("context = %d, want %d", l.Context, tt.wantContext)
+				}
+				if l.Output != tt.wantOutput {
+					t.Errorf("output = %d, want %d", l.Output, tt.wantOutput)
+				}
+			}
+		})
+	}
+}
+
+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/pi.go

@@ -0,0 +1,237 @@
+package config
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/types/model"
+)
+
+// Pi implements Runner and Editor for Pi (Pi Coding Agent) integration
+type Pi struct{}
+
+func (p *Pi) String() string { return "Pi" }
+
+func (p *Pi) Run(model string, args []string) error {
+	if _, err := exec.LookPath("pi"); err != nil {
+		return fmt.Errorf("pi is not installed, install with: npm install -g @mariozechner/pi-coding-agent")
+	}
+
+	// Call Edit() to ensure config is up-to-date before launch
+	models := []string{model}
+	if config, err := loadIntegration("pi"); err == nil && len(config.Models) > 0 {
+		models = config.Models
+	}
+	if err := p.Edit(models); err != nil {
+		return fmt.Errorf("setup failed: %w", err)
+	}
+
+	cmd := exec.Command("pi", args...)
+	cmd.Stdin = os.Stdin
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	return cmd.Run()
+}
+
+func (p *Pi) Paths() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+
+	var paths []string
+	modelsPath := filepath.Join(home, ".pi", "agent", "models.json")
+	if _, err := os.Stat(modelsPath); err == nil {
+		paths = append(paths, modelsPath)
+	}
+	settingsPath := filepath.Join(home, ".pi", "agent", "settings.json")
+	if _, err := os.Stat(settingsPath); err == nil {
+		paths = append(paths, settingsPath)
+	}
+	return paths
+}
+
+func (p *Pi) Edit(models []string) error {
+	if len(models) == 0 {
+		return nil
+	}
+
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return err
+	}
+
+	configPath := filepath.Join(home, ".pi", "agent", "models.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)
+	}
+
+	providers, ok := config["providers"].(map[string]any)
+	if !ok {
+		providers = make(map[string]any)
+	}
+
+	ollama, ok := providers["ollama"].(map[string]any)
+	if !ok {
+		ollama = map[string]any{
+			"baseUrl": envconfig.Host().String() + "/v1",
+			"api":     "openai-completions",
+			"apiKey":  "ollama",
+		}
+	}
+
+	existingModels, ok := ollama["models"].([]any)
+	if !ok {
+		existingModels = make([]any, 0)
+	}
+
+	// Build set of selected models to track which need to be added
+	selectedSet := make(map[string]bool, len(models))
+	for _, m := range models {
+		selectedSet[m] = true
+	}
+
+	// Build new models list:
+	// 1. Keep user-managed models (no _launch marker) - untouched
+	// 2. Keep ollama-managed models (_launch marker) that are still selected
+	// 3. Add new ollama-managed models
+	var newModels []any
+	for _, m := range existingModels {
+		if modelObj, ok := m.(map[string]any); ok {
+			if id, ok := modelObj["id"].(string); ok {
+				// User-managed model (no _launch marker) - always preserve
+				if !isPiOllamaModel(modelObj) {
+					newModels = append(newModels, m)
+				} else if selectedSet[id] {
+					// Ollama-managed and still selected - keep it
+					newModels = append(newModels, m)
+					selectedSet[id] = false
+				}
+			}
+		}
+	}
+
+	// Add newly selected models that weren't already in the list
+	client := api.NewClient(envconfig.Host(), http.DefaultClient)
+	ctx := context.Background()
+	for _, model := range models {
+		if selectedSet[model] {
+			newModels = append(newModels, createConfig(ctx, client, model))
+		}
+	}
+
+	ollama["models"] = newModels
+	providers["ollama"] = ollama
+	config["providers"] = providers
+
+	configData, err := json.MarshalIndent(config, "", "  ")
+	if err != nil {
+		return err
+	}
+	if err := writeWithBackup(configPath, configData); err != nil {
+		return err
+	}
+
+	// Update settings.json with default provider and model
+	settingsPath := filepath.Join(home, ".pi", "agent", "settings.json")
+	settings := make(map[string]any)
+	if data, err := os.ReadFile(settingsPath); err == nil {
+		_ = json.Unmarshal(data, &settings)
+	}
+
+	settings["defaultProvider"] = "ollama"
+	settings["defaultModel"] = models[0]
+
+	settingsData, err := json.MarshalIndent(settings, "", "  ")
+	if err != nil {
+		return err
+	}
+	return writeWithBackup(settingsPath, settingsData)
+}
+
+func (p *Pi) Models() []string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return nil
+	}
+
+	configPath := filepath.Join(home, ".pi", "agent", "models.json")
+	config, err := readJSONFile(configPath)
+	if err != nil {
+		return nil
+	}
+
+	providers, _ := config["providers"].(map[string]any)
+	ollama, _ := providers["ollama"].(map[string]any)
+	models, _ := ollama["models"].([]any)
+
+	var result []string
+	for _, m := range models {
+		if modelObj, ok := m.(map[string]any); ok {
+			if id, ok := modelObj["id"].(string); ok {
+				result = append(result, id)
+			}
+		}
+	}
+	slices.Sort(result)
+	return result
+}
+
+// isPiOllamaModel reports whether a model config entry is managed by ollama launch
+func isPiOllamaModel(cfg map[string]any) bool {
+	if v, ok := cfg["_launch"].(bool); ok && v {
+		return true
+	}
+	return false
+}
+
+// createConfig builds Pi model config with capability detection
+func createConfig(ctx context.Context, client *api.Client, modelID string) map[string]any {
+	cfg := map[string]any{
+		"id":      modelID,
+		"_launch": true,
+	}
+
+	resp, err := client.Show(ctx, &api.ShowRequest{Model: modelID})
+	if err != nil {
+		return cfg
+	}
+
+	// Set input types based on vision capability
+	if slices.Contains(resp.Capabilities, model.CapabilityVision) {
+		cfg["input"] = []string{"text", "image"}
+	} else {
+		cfg["input"] = []string{"text"}
+	}
+
+	// Set reasoning based on thinking capability
+	if slices.Contains(resp.Capabilities, model.CapabilityThinking) {
+		cfg["reasoning"] = true
+	}
+
+	// Extract context window from ModelInfo
+	for key, val := range resp.ModelInfo {
+		if strings.HasSuffix(key, ".context_length") {
+			if ctxLen, ok := val.(float64); ok && ctxLen > 0 {
+				cfg["contextWindow"] = int(ctxLen)
+			}
+			break
+		}
+	}
+
+	return cfg
+}

cmd/config/pi_test.go

@@ -0,0 +1,830 @@
+package config
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/types/model"
+)
+
+func TestPiIntegration(t *testing.T) {
+	pi := &Pi{}
+
+	t.Run("String", func(t *testing.T) {
+		if got := pi.String(); got != "Pi" {
+			t.Errorf("String() = %q, want %q", got, "Pi")
+		}
+	})
+
+	t.Run("implements Runner", func(t *testing.T) {
+		var _ Runner = pi
+	})
+
+	t.Run("implements Editor", func(t *testing.T) {
+		var _ Editor = pi
+	})
+}
+
+func TestPiPaths(t *testing.T) {
+	pi := &Pi{}
+
+	t.Run("returns empty when no config exists", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		paths := pi.Paths()
+		if len(paths) != 0 {
+			t.Errorf("Paths() = %v, want empty", paths)
+		}
+	})
+
+	t.Run("returns path when config exists", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		configDir := filepath.Join(tmpDir, ".pi", "agent")
+		if err := os.MkdirAll(configDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		configPath := filepath.Join(configDir, "models.json")
+		if err := os.WriteFile(configPath, []byte("{}"), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		paths := pi.Paths()
+		if len(paths) != 1 || paths[0] != configPath {
+			t.Errorf("Paths() = %v, want [%s]", paths, configPath)
+		}
+	})
+}
+
+func TestPiEdit(t *testing.T) {
+	// Mock Ollama server for createConfig calls during Edit
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/api/show" {
+			fmt.Fprintf(w, `{"capabilities":[],"model_info":{}}`)
+			return
+		}
+		w.WriteHeader(http.StatusNotFound)
+	}))
+	defer srv.Close()
+	t.Setenv("OLLAMA_HOST", srv.URL)
+
+	pi := &Pi{}
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	configDir := filepath.Join(tmpDir, ".pi", "agent")
+	configPath := filepath.Join(configDir, "models.json")
+
+	cleanup := func() {
+		os.RemoveAll(configDir)
+	}
+
+	readConfig := func() map[string]any {
+		data, _ := os.ReadFile(configPath)
+		var cfg map[string]any
+		json.Unmarshal(data, &cfg)
+		return cfg
+	}
+
+	t.Run("returns nil for empty models", func(t *testing.T) {
+		if err := pi.Edit([]string{}); err != nil {
+			t.Errorf("Edit([]) error = %v, want nil", err)
+		}
+	})
+
+	t.Run("creates config with models", func(t *testing.T) {
+		cleanup()
+
+		models := []string{"llama3.2", "qwen3:8b"}
+		if err := pi.Edit(models); err != nil {
+			t.Fatalf("Edit() error = %v", err)
+		}
+
+		cfg := readConfig()
+
+		providers, ok := cfg["providers"].(map[string]any)
+		if !ok {
+			t.Error("Config missing providers")
+		}
+
+		ollama, ok := providers["ollama"].(map[string]any)
+		if !ok {
+			t.Error("Providers missing ollama")
+		}
+
+		modelsArray, ok := ollama["models"].([]any)
+		if !ok || len(modelsArray) != 2 {
+			t.Errorf("Expected 2 models, got %v", modelsArray)
+		}
+
+		if ollama["baseUrl"] == nil {
+			t.Error("Missing baseUrl")
+		}
+		if ollama["api"] != "openai-completions" {
+			t.Errorf("Expected api=openai-completions, got %v", ollama["api"])
+		}
+		if ollama["apiKey"] != "ollama" {
+			t.Errorf("Expected apiKey=ollama, got %v", ollama["apiKey"])
+		}
+	})
+
+	t.Run("updates existing config preserving ollama provider settings", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+
+		existingConfig := `{
+			"providers": {
+				"ollama": {
+					"baseUrl": "http://custom:8080/v1",
+					"api": "custom-api",
+					"apiKey": "custom-key",
+					"models": [
+						{"id": "old-model", "_launch": true}
+					]
+				}
+			}
+		}`
+		if err := os.WriteFile(configPath, []byte(existingConfig), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		models := []string{"new-model"}
+		if err := pi.Edit(models); err != nil {
+			t.Fatalf("Edit() error = %v", err)
+		}
+
+		cfg := readConfig()
+		providers := cfg["providers"].(map[string]any)
+		ollama := providers["ollama"].(map[string]any)
+
+		if ollama["baseUrl"] != "http://custom:8080/v1" {
+			t.Errorf("Custom baseUrl not preserved, got %v", ollama["baseUrl"])
+		}
+		if ollama["api"] != "custom-api" {
+			t.Errorf("Custom api not preserved, got %v", ollama["api"])
+		}
+		if ollama["apiKey"] != "custom-key" {
+			t.Errorf("Custom apiKey not preserved, got %v", ollama["apiKey"])
+		}
+
+		modelsArray := ollama["models"].([]any)
+		if len(modelsArray) != 1 {
+			t.Errorf("Expected 1 model after update, got %d", len(modelsArray))
+		} else {
+			modelEntry := modelsArray[0].(map[string]any)
+			if modelEntry["id"] != "new-model" {
+				t.Errorf("Expected new-model, got %v", modelEntry["id"])
+			}
+			// Verify _launch marker is present
+			if modelEntry["_launch"] != true {
+				t.Errorf("Expected _launch marker to be true")
+			}
+		}
+	})
+
+	t.Run("replaces old models with new ones", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+
+		// Old models must have _launch marker to be managed by us
+		existingConfig := `{
+			"providers": {
+				"ollama": {
+					"baseUrl": "http://localhost:11434/v1",
+					"api": "openai-completions",
+					"apiKey": "ollama",
+					"models": [
+						{"id": "old-model-1", "_launch": true},
+						{"id": "old-model-2", "_launch": true}
+					]
+				}
+			}
+		}`
+		if err := os.WriteFile(configPath, []byte(existingConfig), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		newModels := []string{"new-model-1", "new-model-2"}
+		if err := pi.Edit(newModels); err != nil {
+			t.Fatalf("Edit() error = %v", err)
+		}
+
+		cfg := readConfig()
+		providers := cfg["providers"].(map[string]any)
+		ollama := providers["ollama"].(map[string]any)
+		modelsArray := ollama["models"].([]any)
+
+		if len(modelsArray) != 2 {
+			t.Errorf("Expected 2 models, got %d", len(modelsArray))
+		}
+
+		modelIDs := make(map[string]bool)
+		for _, m := range modelsArray {
+			modelObj := m.(map[string]any)
+			id := modelObj["id"].(string)
+			modelIDs[id] = true
+		}
+
+		if !modelIDs["new-model-1"] || !modelIDs["new-model-2"] {
+			t.Errorf("Expected new models, got %v", modelIDs)
+		}
+		if modelIDs["old-model-1"] || modelIDs["old-model-2"] {
+			t.Errorf("Old models should have been removed, got %v", modelIDs)
+		}
+	})
+
+	t.Run("handles partial overlap in model list", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+
+		// Models must have _launch marker to be managed
+		existingConfig := `{
+			"providers": {
+				"ollama": {
+					"baseUrl": "http://localhost:11434/v1",
+					"api": "openai-completions",
+					"apiKey": "ollama",
+					"models": [
+						{"id": "keep-model", "_launch": true},
+						{"id": "remove-model", "_launch": true}
+					]
+				}
+			}
+		}`
+		if err := os.WriteFile(configPath, []byte(existingConfig), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		newModels := []string{"keep-model", "add-model"}
+		if err := pi.Edit(newModels); err != nil {
+			t.Fatalf("Edit() error = %v", err)
+		}
+
+		cfg := readConfig()
+		providers := cfg["providers"].(map[string]any)
+		ollama := providers["ollama"].(map[string]any)
+		modelsArray := ollama["models"].([]any)
+
+		if len(modelsArray) != 2 {
+			t.Errorf("Expected 2 models, got %d", len(modelsArray))
+		}
+
+		modelIDs := make(map[string]bool)
+		for _, m := range modelsArray {
+			modelObj := m.(map[string]any)
+			id := modelObj["id"].(string)
+			modelIDs[id] = true
+		}
+
+		if !modelIDs["keep-model"] || !modelIDs["add-model"] {
+			t.Errorf("Expected keep-model and add-model, got %v", modelIDs)
+		}
+		if modelIDs["remove-model"] {
+			t.Errorf("remove-model should have been removed")
+		}
+	})
+
+	t.Run("handles corrupt config gracefully", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+
+		if err := os.WriteFile(configPath, []byte("{invalid json}"), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		models := []string{"test-model"}
+		if err := pi.Edit(models); err != nil {
+			t.Fatalf("Edit() should not fail with corrupt config, got %v", err)
+		}
+
+		data, err := os.ReadFile(configPath)
+		if err != nil {
+			t.Fatalf("Failed to read config: %v", err)
+		}
+
+		var cfg map[string]any
+		if err := json.Unmarshal(data, &cfg); err != nil {
+			t.Fatalf("Config should be valid after Edit, got parse error: %v", err)
+		}
+
+		providers := cfg["providers"].(map[string]any)
+		ollama := providers["ollama"].(map[string]any)
+		modelsArray := ollama["models"].([]any)
+
+		if len(modelsArray) != 1 {
+			t.Errorf("Expected 1 model, got %d", len(modelsArray))
+		}
+	})
+
+	// CRITICAL SAFETY TEST: verifies we don't stomp on user configs
+	t.Run("preserves user-managed models without _launch marker", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+
+		// User has manually configured models in ollama provider (no _launch marker)
+		existingConfig := `{
+			"providers": {
+				"ollama": {
+					"baseUrl": "http://localhost:11434/v1",
+					"api": "openai-completions",
+					"apiKey": "ollama",
+					"models": [
+						{"id": "user-model-1"},
+						{"id": "user-model-2", "customField": "preserved"},
+						{"id": "ollama-managed", "_launch": true}
+					]
+				}
+			}
+		}`
+		if err := os.WriteFile(configPath, []byte(existingConfig), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		// Add a new ollama-managed model
+		newModels := []string{"new-ollama-model"}
+		if err := pi.Edit(newModels); err != nil {
+			t.Fatalf("Edit() error = %v", err)
+		}
+
+		cfg := readConfig()
+		providers := cfg["providers"].(map[string]any)
+		ollama := providers["ollama"].(map[string]any)
+		modelsArray := ollama["models"].([]any)
+
+		// Should have: new-ollama-model (managed) + 2 user models (preserved)
+		if len(modelsArray) != 3 {
+			t.Errorf("Expected 3 models (1 new managed + 2 preserved user models), got %d", len(modelsArray))
+		}
+
+		modelIDs := make(map[string]map[string]any)
+		for _, m := range modelsArray {
+			modelObj := m.(map[string]any)
+			id := modelObj["id"].(string)
+			modelIDs[id] = modelObj
+		}
+
+		// Verify new model has _launch marker
+		if m, ok := modelIDs["new-ollama-model"]; !ok {
+			t.Errorf("new-ollama-model should be present")
+		} else if m["_launch"] != true {
+			t.Errorf("new-ollama-model should have _launch marker")
+		}
+
+		// Verify user models are preserved
+		if _, ok := modelIDs["user-model-1"]; !ok {
+			t.Errorf("user-model-1 should be preserved")
+		}
+		if _, ok := modelIDs["user-model-2"]; !ok {
+			t.Errorf("user-model-2 should be preserved")
+		} else if modelIDs["user-model-2"]["customField"] != "preserved" {
+			t.Errorf("user-model-2 customField should be preserved")
+		}
+
+		// Verify old ollama-managed model is removed (not in new list)
+		if _, ok := modelIDs["ollama-managed"]; ok {
+			t.Errorf("ollama-managed should be removed (old ollama model not in new selection)")
+		}
+	})
+
+	t.Run("updates settings.json with default provider and model", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+
+		// Create existing settings with other fields
+		settingsPath := filepath.Join(configDir, "settings.json")
+		existingSettings := `{
+			"theme": "dark",
+			"customSetting": "value",
+			"defaultProvider": "anthropic",
+			"defaultModel": "claude-3"
+		}`
+		if err := os.WriteFile(settingsPath, []byte(existingSettings), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		models := []string{"llama3.2"}
+		if err := pi.Edit(models); err != nil {
+			t.Fatalf("Edit() error = %v", err)
+		}
+
+		data, err := os.ReadFile(settingsPath)
+		if err != nil {
+			t.Fatalf("Failed to read settings: %v", err)
+		}
+
+		var settings map[string]any
+		if err := json.Unmarshal(data, &settings); err != nil {
+			t.Fatalf("Failed to parse settings: %v", err)
+		}
+
+		// Verify defaultProvider is set to ollama
+		if settings["defaultProvider"] != "ollama" {
+			t.Errorf("defaultProvider = %v, want ollama", settings["defaultProvider"])
+		}
+
+		// Verify defaultModel is set to first model
+		if settings["defaultModel"] != "llama3.2" {
+			t.Errorf("defaultModel = %v, want llama3.2", settings["defaultModel"])
+		}
+
+		// Verify other fields are preserved
+		if settings["theme"] != "dark" {
+			t.Errorf("theme = %v, want dark (preserved)", settings["theme"])
+		}
+		if settings["customSetting"] != "value" {
+			t.Errorf("customSetting = %v, want value (preserved)", settings["customSetting"])
+		}
+	})
+
+	t.Run("creates settings.json if it does not exist", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+
+		models := []string{"qwen3:8b"}
+		if err := pi.Edit(models); err != nil {
+			t.Fatalf("Edit() error = %v", err)
+		}
+
+		settingsPath := filepath.Join(configDir, "settings.json")
+		data, err := os.ReadFile(settingsPath)
+		if err != nil {
+			t.Fatalf("settings.json should be created: %v", err)
+		}
+
+		var settings map[string]any
+		if err := json.Unmarshal(data, &settings); err != nil {
+			t.Fatalf("Failed to parse settings: %v", err)
+		}
+
+		if settings["defaultProvider"] != "ollama" {
+			t.Errorf("defaultProvider = %v, want ollama", settings["defaultProvider"])
+		}
+		if settings["defaultModel"] != "qwen3:8b" {
+			t.Errorf("defaultModel = %v, want qwen3:8b", settings["defaultModel"])
+		}
+	})
+
+	t.Run("handles corrupt settings.json gracefully", func(t *testing.T) {
+		cleanup()
+		os.MkdirAll(configDir, 0o755)
+
+		// Create corrupt settings
+		settingsPath := filepath.Join(configDir, "settings.json")
+		if err := os.WriteFile(settingsPath, []byte("{invalid"), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		models := []string{"test-model"}
+		if err := pi.Edit(models); err != nil {
+			t.Fatalf("Edit() should not fail with corrupt settings, got %v", err)
+		}
+
+		data, err := os.ReadFile(settingsPath)
+		if err != nil {
+			t.Fatalf("Failed to read settings: %v", err)
+		}
+
+		var settings map[string]any
+		if err := json.Unmarshal(data, &settings); err != nil {
+			t.Fatalf("settings.json should be valid after Edit, got parse error: %v", err)
+		}
+
+		if settings["defaultProvider"] != "ollama" {
+			t.Errorf("defaultProvider = %v, want ollama", settings["defaultProvider"])
+		}
+		if settings["defaultModel"] != "test-model" {
+			t.Errorf("defaultModel = %v, want test-model", settings["defaultModel"])
+		}
+	})
+}
+
+func TestPiModels(t *testing.T) {
+	pi := &Pi{}
+
+	t.Run("returns nil when no config exists", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		models := pi.Models()
+		if models != nil {
+			t.Errorf("Models() = %v, want nil", models)
+		}
+	})
+
+	t.Run("returns models from config", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		configDir := filepath.Join(tmpDir, ".pi", "agent")
+		if err := os.MkdirAll(configDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		config := `{
+			"providers": {
+				"ollama": {
+					"models": [
+						{"id": "llama3.2"},
+						{"id": "qwen3:8b"}
+					]
+				}
+			}
+		}`
+		configPath := filepath.Join(configDir, "models.json")
+		if err := os.WriteFile(configPath, []byte(config), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		models := pi.Models()
+		if len(models) != 2 {
+			t.Errorf("Models() returned %d models, want 2", len(models))
+		}
+		if models[0] != "llama3.2" || models[1] != "qwen3:8b" {
+			t.Errorf("Models() = %v, want [llama3.2 qwen3:8b] (sorted)", models)
+		}
+	})
+
+	t.Run("returns sorted models", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		configDir := filepath.Join(tmpDir, ".pi", "agent")
+		if err := os.MkdirAll(configDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		config := `{
+			"providers": {
+				"ollama": {
+					"models": [
+						{"id": "z-model"},
+						{"id": "a-model"},
+						{"id": "m-model"}
+					]
+				}
+			}
+		}`
+		configPath := filepath.Join(configDir, "models.json")
+		if err := os.WriteFile(configPath, []byte(config), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		models := pi.Models()
+		if models[0] != "a-model" || models[1] != "m-model" || models[2] != "z-model" {
+			t.Errorf("Models() = %v, want [a-model m-model z-model] (sorted)", models)
+		}
+	})
+
+	t.Run("returns nil when models array is missing", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		configDir := filepath.Join(tmpDir, ".pi", "agent")
+		if err := os.MkdirAll(configDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		config := `{
+			"providers": {
+				"ollama": {}
+			}
+		}`
+		configPath := filepath.Join(configDir, "models.json")
+		if err := os.WriteFile(configPath, []byte(config), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		models := pi.Models()
+		if models != nil {
+			t.Errorf("Models() = %v, want nil when models array is missing", models)
+		}
+	})
+
+	t.Run("handles corrupt config gracefully", func(t *testing.T) {
+		tmpDir := t.TempDir()
+		setTestHome(t, tmpDir)
+
+		configDir := filepath.Join(tmpDir, ".pi", "agent")
+		if err := os.MkdirAll(configDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		configPath := filepath.Join(configDir, "models.json")
+		if err := os.WriteFile(configPath, []byte("{invalid json}"), 0o644); err != nil {
+			t.Fatal(err)
+		}
+
+		models := pi.Models()
+		if models != nil {
+			t.Errorf("Models() = %v, want nil for corrupt config", models)
+		}
+	})
+}
+
+func TestIsPiOllamaModel(t *testing.T) {
+	tests := []struct {
+		name string
+		cfg  map[string]any
+		want bool
+	}{
+		{"with _launch true", map[string]any{"id": "m", "_launch": true}, true},
+		{"with _launch false", map[string]any{"id": "m", "_launch": false}, false},
+		{"without _launch", map[string]any{"id": "m"}, false},
+		{"with _launch non-bool", map[string]any{"id": "m", "_launch": "yes"}, false},
+		{"empty map", map[string]any{}, false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := isPiOllamaModel(tt.cfg); got != tt.want {
+				t.Errorf("isPiOllamaModel(%v) = %v, want %v", tt.cfg, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestCreateConfig(t *testing.T) {
+	t.Run("sets vision input when model has vision capability", func(t *testing.T) {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			if r.URL.Path == "/api/show" {
+				fmt.Fprintf(w, `{"capabilities":["vision"],"model_info":{}}`)
+				return
+			}
+			w.WriteHeader(http.StatusNotFound)
+		}))
+		defer srv.Close()
+
+		u, _ := url.Parse(srv.URL)
+		client := api.NewClient(u, srv.Client())
+
+		cfg := createConfig(context.Background(), client, "llava:7b")
+
+		if cfg["id"] != "llava:7b" {
+			t.Errorf("id = %v, want llava:7b", cfg["id"])
+		}
+		if cfg["_launch"] != true {
+			t.Error("expected _launch = true")
+		}
+		input, ok := cfg["input"].([]string)
+		if !ok || len(input) != 2 || input[0] != "text" || input[1] != "image" {
+			t.Errorf("input = %v, want [text image]", cfg["input"])
+		}
+	})
+
+	t.Run("sets text-only input when model lacks vision", func(t *testing.T) {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			if r.URL.Path == "/api/show" {
+				fmt.Fprintf(w, `{"capabilities":["completion"],"model_info":{}}`)
+				return
+			}
+			w.WriteHeader(http.StatusNotFound)
+		}))
+		defer srv.Close()
+
+		u, _ := url.Parse(srv.URL)
+		client := api.NewClient(u, srv.Client())
+
+		cfg := createConfig(context.Background(), client, "llama3.2")
+
+		input, ok := cfg["input"].([]string)
+		if !ok || len(input) != 1 || input[0] != "text" {
+			t.Errorf("input = %v, want [text]", cfg["input"])
+		}
+		if _, ok := cfg["reasoning"]; ok {
+			t.Error("reasoning should not be set for non-thinking model")
+		}
+	})
+
+	t.Run("sets reasoning when model has thinking capability", func(t *testing.T) {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			if r.URL.Path == "/api/show" {
+				fmt.Fprintf(w, `{"capabilities":["thinking"],"model_info":{}}`)
+				return
+			}
+			w.WriteHeader(http.StatusNotFound)
+		}))
+		defer srv.Close()
+
+		u, _ := url.Parse(srv.URL)
+		client := api.NewClient(u, srv.Client())
+
+		cfg := createConfig(context.Background(), client, "qwq")
+
+		if cfg["reasoning"] != true {
+			t.Error("expected reasoning = true for thinking model")
+		}
+	})
+
+	t.Run("extracts context window from model info", func(t *testing.T) {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			if r.URL.Path == "/api/show" {
+				fmt.Fprintf(w, `{"capabilities":[],"model_info":{"llama.context_length":131072}}`)
+				return
+			}
+			w.WriteHeader(http.StatusNotFound)
+		}))
+		defer srv.Close()
+
+		u, _ := url.Parse(srv.URL)
+		client := api.NewClient(u, srv.Client())
+
+		cfg := createConfig(context.Background(), client, "llama3.2")
+
+		if cfg["contextWindow"] != 131072 {
+			t.Errorf("contextWindow = %v, want 131072", cfg["contextWindow"])
+		}
+	})
+
+	t.Run("handles all capabilities together", func(t *testing.T) {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			if r.URL.Path == "/api/show" {
+				fmt.Fprintf(w, `{"capabilities":["vision","thinking"],"model_info":{"qwen3.context_length":32768}}`)
+				return
+			}
+			w.WriteHeader(http.StatusNotFound)
+		}))
+		defer srv.Close()
+
+		u, _ := url.Parse(srv.URL)
+		client := api.NewClient(u, srv.Client())
+
+		cfg := createConfig(context.Background(), client, "qwen3-vision")
+
+		input := cfg["input"].([]string)
+		if len(input) != 2 || input[0] != "text" || input[1] != "image" {
+			t.Errorf("input = %v, want [text image]", input)
+		}
+		if cfg["reasoning"] != true {
+			t.Error("expected reasoning = true")
+		}
+		if cfg["contextWindow"] != 32768 {
+			t.Errorf("contextWindow = %v, want 32768", cfg["contextWindow"])
+		}
+	})
+
+	t.Run("returns minimal config when show fails", func(t *testing.T) {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			w.WriteHeader(http.StatusNotFound)
+			fmt.Fprintf(w, `{"error":"model not found"}`)
+		}))
+		defer srv.Close()
+
+		u, _ := url.Parse(srv.URL)
+		client := api.NewClient(u, srv.Client())
+
+		cfg := createConfig(context.Background(), client, "missing-model")
+
+		if cfg["id"] != "missing-model" {
+			t.Errorf("id = %v, want missing-model", cfg["id"])
+		}
+		if cfg["_launch"] != true {
+			t.Error("expected _launch = true")
+		}
+		// Should not have capability fields
+		if _, ok := cfg["input"]; ok {
+			t.Error("input should not be set when show fails")
+		}
+		if _, ok := cfg["reasoning"]; ok {
+			t.Error("reasoning should not be set when show fails")
+		}
+		if _, ok := cfg["contextWindow"]; ok {
+			t.Error("contextWindow should not be set when show fails")
+		}
+	})
+
+	t.Run("skips zero context length", func(t *testing.T) {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			if r.URL.Path == "/api/show" {
+				fmt.Fprintf(w, `{"capabilities":[],"model_info":{"llama.context_length":0}}`)
+				return
+			}
+			w.WriteHeader(http.StatusNotFound)
+		}))
+		defer srv.Close()
+
+		u, _ := url.Parse(srv.URL)
+		client := api.NewClient(u, srv.Client())
+
+		cfg := createConfig(context.Background(), client, "test-model")
+
+		if _, ok := cfg["contextWindow"]; ok {
+			t.Error("contextWindow should not be set for zero value")
+		}
+	})
+}
+
+// Ensure Capability constants used in createConfig match expected values
+func TestPiCapabilityConstants(t *testing.T) {
+	if model.CapabilityVision != "vision" {
+		t.Errorf("CapabilityVision = %q, want %q", model.CapabilityVision, "vision")
+	}
+	if model.CapabilityThinking != "thinking" {
+		t.Errorf("CapabilityThinking = %q, want %q", model.CapabilityThinking, "thinking")
+	}
+}

cmd/config/selector.go

@@ -0,0 +1,58 @@
+package config
+
+import (
+	"errors"
+	"fmt"
+	"os"
+
+	"golang.org/x/term"
+)
+
+// ANSI escape sequences for terminal formatting.
+const (
+	ansiBold  = "\033[1m"
+	ansiReset = "\033[0m"
+	ansiGray  = "\033[37m"
+	ansiGreen = "\033[32m"
+)
+
+// ErrCancelled is returned when the user cancels a selection.
+var ErrCancelled = errors.New("cancelled")
+
+// errCancelled is kept as an alias for backward compatibility within the package.
+var errCancelled = ErrCancelled
+
+// DefaultConfirmPrompt provides a TUI-based confirmation prompt.
+// When set, confirmPrompt delegates to it instead of using raw terminal I/O.
+var DefaultConfirmPrompt func(prompt string) (bool, error)
+
+func confirmPrompt(prompt string) (bool, error) {
+	if DefaultConfirmPrompt != nil {
+		return DefaultConfirmPrompt(prompt)
+	}
+
+	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 (\033[1my\033[0m/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
+		}
+	}
+}

cmd/config/selector_test.go

@@ -0,0 +1,19 @@
+package config
+
+import (
+	"testing"
+)
+
+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())
+		}
+	})
+}

cmd/interactive.go

@@ -159,6 +159,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			sb.WriteString(before)
 			if !ok {
 				fmt.Fprintln(&sb)
+				scanner.Prompt.UseAlt = true
 				continue
 			}
 

cmd/start_darwin.go

@@ -10,19 +10,21 @@ import (
 	"github.com/ollama/ollama/api"
 )
 
+var errNotRunning = errors.New("could not connect to ollama server, run 'ollama serve' to start it")
+
 func startApp(ctx context.Context, client *api.Client) error {
 	exe, err := os.Executable()
 	if err != nil {
-		return err
+		return errNotRunning
 	}
 	link, err := os.Readlink(exe)
 	if err != nil {
-		return err
+		return errNotRunning
 	}
 	r := regexp.MustCompile(`^.*/Ollama\s?\d*.app`)
 	m := r.FindStringSubmatch(link)
 	if len(m) != 1 {
-		return errors.New("could not find ollama app")
+		return errNotRunning
 	}
 	if err := exec.Command("/usr/bin/open", "-j", "-a", m[0], "--args", "--fast-startup").Run(); err != nil {
 		return err

cmd/tui/confirm.go

@@ -0,0 +1,109 @@
+package tui
+
+import (
+	"fmt"
+
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+)
+
+var (
+	confirmActiveStyle = lipgloss.NewStyle().
+				Bold(true).
+				Background(lipgloss.AdaptiveColor{Light: "254", Dark: "236"})
+
+	confirmInactiveStyle = lipgloss.NewStyle().
+				Foreground(lipgloss.AdaptiveColor{Light: "242", Dark: "246"})
+)
+
+type confirmModel struct {
+	prompt    string
+	yes       bool
+	confirmed bool
+	cancelled bool
+	width     int
+}
+
+func (m confirmModel) Init() tea.Cmd {
+	return nil
+}
+
+func (m confirmModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
+	switch msg := msg.(type) {
+	case tea.WindowSizeMsg:
+		wasSet := m.width > 0
+		m.width = msg.Width
+		if wasSet {
+			return m, tea.EnterAltScreen
+		}
+		return m, nil
+
+	case tea.KeyMsg:
+		switch msg.String() {
+		case "ctrl+c", "esc", "n":
+			m.cancelled = true
+			return m, tea.Quit
+		case "y":
+			m.yes = true
+			m.confirmed = true
+			return m, tea.Quit
+		case "enter":
+			m.confirmed = true
+			return m, tea.Quit
+		case "left", "h":
+			m.yes = true
+		case "right", "l":
+			m.yes = false
+		case "tab":
+			m.yes = !m.yes
+		}
+	}
+
+	return m, nil
+}
+
+func (m confirmModel) View() string {
+	if m.confirmed || m.cancelled {
+		return ""
+	}
+
+	var yesBtn, noBtn string
+	if m.yes {
+		yesBtn = confirmActiveStyle.Render(" Yes ")
+		noBtn = confirmInactiveStyle.Render(" No ")
+	} else {
+		yesBtn = confirmInactiveStyle.Render(" Yes ")
+		noBtn = confirmActiveStyle.Render(" No ")
+	}
+
+	s := selectorTitleStyle.Render(m.prompt) + "\n\n"
+	s += "  " + yesBtn + "  " + noBtn + "\n\n"
+	s += selectorHelpStyle.Render("←/→ navigate • enter confirm • esc cancel")
+
+	if m.width > 0 {
+		return lipgloss.NewStyle().MaxWidth(m.width).Render(s)
+	}
+	return s
+}
+
+// RunConfirm shows a bubbletea yes/no confirmation prompt.
+// Returns true if the user confirmed, false if cancelled.
+func RunConfirm(prompt string) (bool, error) {
+	m := confirmModel{
+		prompt: prompt,
+		yes:    true, // default to yes
+	}
+
+	p := tea.NewProgram(m)
+	finalModel, err := p.Run()
+	if err != nil {
+		return false, fmt.Errorf("error running confirm: %w", err)
+	}
+
+	fm := finalModel.(confirmModel)
+	if fm.cancelled {
+		return false, ErrCancelled
+	}
+
+	return fm.yes, nil
+}

cmd/tui/confirm_test.go

@@ -0,0 +1,208 @@
+package tui
+
+import (
+	"strings"
+	"testing"
+
+	tea "github.com/charmbracelet/bubbletea"
+)
+
+func TestConfirmModel_DefaultsToYes(t *testing.T) {
+	m := confirmModel{prompt: "Download test?", yes: true}
+	if !m.yes {
+		t.Error("should default to yes")
+	}
+}
+
+func TestConfirmModel_View_ContainsPrompt(t *testing.T) {
+	m := confirmModel{prompt: "Download qwen3:8b?", yes: true}
+	got := m.View()
+	if !strings.Contains(got, "Download qwen3:8b?") {
+		t.Error("should contain the prompt text")
+	}
+}
+
+func TestConfirmModel_View_ContainsButtons(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: true}
+	got := m.View()
+	if !strings.Contains(got, "Yes") {
+		t.Error("should contain Yes button")
+	}
+	if !strings.Contains(got, "No") {
+		t.Error("should contain No button")
+	}
+}
+
+func TestConfirmModel_View_ContainsHelp(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: true}
+	got := m.View()
+	if !strings.Contains(got, "enter confirm") {
+		t.Error("should contain help text")
+	}
+}
+
+func TestConfirmModel_View_ClearsAfterConfirm(t *testing.T) {
+	m := confirmModel{prompt: "Download?", confirmed: true}
+	if m.View() != "" {
+		t.Error("View should return empty string after confirmation")
+	}
+}
+
+func TestConfirmModel_View_ClearsAfterCancel(t *testing.T) {
+	m := confirmModel{prompt: "Download?", cancelled: true}
+	if m.View() != "" {
+		t.Error("View should return empty string after cancellation")
+	}
+}
+
+func TestConfirmModel_EnterConfirmsYes(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: true}
+	updated, cmd := m.Update(tea.KeyMsg{Type: tea.KeyEnter})
+	fm := updated.(confirmModel)
+	if !fm.confirmed {
+		t.Error("enter should set confirmed=true")
+	}
+	if !fm.yes {
+		t.Error("enter with yes selected should keep yes=true")
+	}
+	if cmd == nil {
+		t.Error("enter should return tea.Quit")
+	}
+}
+
+func TestConfirmModel_EnterConfirmsNo(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: false}
+	updated, cmd := m.Update(tea.KeyMsg{Type: tea.KeyEnter})
+	fm := updated.(confirmModel)
+	if !fm.confirmed {
+		t.Error("enter should set confirmed=true")
+	}
+	if fm.yes {
+		t.Error("enter with no selected should keep yes=false")
+	}
+	if cmd == nil {
+		t.Error("enter should return tea.Quit")
+	}
+}
+
+func TestConfirmModel_EscCancels(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: true}
+	updated, cmd := m.Update(tea.KeyMsg{Type: tea.KeyEsc})
+	fm := updated.(confirmModel)
+	if !fm.cancelled {
+		t.Error("esc should set cancelled=true")
+	}
+	if cmd == nil {
+		t.Error("esc should return tea.Quit")
+	}
+}
+
+func TestConfirmModel_CtrlCCancels(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: true}
+	updated, cmd := m.Update(tea.KeyMsg{Type: tea.KeyCtrlC})
+	fm := updated.(confirmModel)
+	if !fm.cancelled {
+		t.Error("ctrl+c should set cancelled=true")
+	}
+	if cmd == nil {
+		t.Error("ctrl+c should return tea.Quit")
+	}
+}
+
+func TestConfirmModel_NCancels(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: true}
+	updated, cmd := m.Update(tea.KeyMsg{Type: tea.KeyRunes, Runes: []rune{'n'}})
+	fm := updated.(confirmModel)
+	if !fm.cancelled {
+		t.Error("'n' should set cancelled=true")
+	}
+	if cmd == nil {
+		t.Error("'n' should return tea.Quit")
+	}
+}
+
+func TestConfirmModel_YConfirmsYes(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: false}
+	updated, cmd := m.Update(tea.KeyMsg{Type: tea.KeyRunes, Runes: []rune{'y'}})
+	fm := updated.(confirmModel)
+	if !fm.confirmed {
+		t.Error("'y' should set confirmed=true")
+	}
+	if !fm.yes {
+		t.Error("'y' should set yes=true")
+	}
+	if cmd == nil {
+		t.Error("'y' should return tea.Quit")
+	}
+}
+
+func TestConfirmModel_ArrowKeysNavigate(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: true}
+
+	// Right moves to No
+	updated, _ := m.Update(tea.KeyMsg{Type: tea.KeyRunes, Runes: []rune{'l'}})
+	fm := updated.(confirmModel)
+	if fm.yes {
+		t.Error("right/l should move to No")
+	}
+	if fm.confirmed || fm.cancelled {
+		t.Error("navigation should not confirm or cancel")
+	}
+
+	// Left moves back to Yes
+	updated, _ = fm.Update(tea.KeyMsg{Type: tea.KeyRunes, Runes: []rune{'h'}})
+	fm = updated.(confirmModel)
+	if !fm.yes {
+		t.Error("left/h should move to Yes")
+	}
+}
+
+func TestConfirmModel_TabToggles(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: true}
+
+	updated, _ := m.Update(tea.KeyMsg{Type: tea.KeyTab})
+	fm := updated.(confirmModel)
+	if fm.yes {
+		t.Error("tab should toggle from Yes to No")
+	}
+
+	updated, _ = fm.Update(tea.KeyMsg{Type: tea.KeyTab})
+	fm = updated.(confirmModel)
+	if !fm.yes {
+		t.Error("tab should toggle from No to Yes")
+	}
+}
+
+func TestConfirmModel_WindowSizeUpdatesWidth(t *testing.T) {
+	m := confirmModel{prompt: "Download?"}
+	updated, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 40})
+	fm := updated.(confirmModel)
+	if fm.width != 100 {
+		t.Errorf("expected width 100, got %d", fm.width)
+	}
+}
+
+func TestConfirmModel_ResizeEntersAltScreen(t *testing.T) {
+	m := confirmModel{prompt: "Download?", width: 80}
+	_, cmd := m.Update(tea.WindowSizeMsg{Width: 100, Height: 40})
+	if cmd == nil {
+		t.Error("resize (width already set) should return a command")
+	}
+}
+
+func TestConfirmModel_InitialWindowSizeNoAltScreen(t *testing.T) {
+	m := confirmModel{prompt: "Download?"}
+	_, cmd := m.Update(tea.WindowSizeMsg{Width: 80, Height: 40})
+	if cmd != nil {
+		t.Error("initial WindowSizeMsg should not return a command")
+	}
+}
+
+func TestConfirmModel_ViewMaxWidth(t *testing.T) {
+	m := confirmModel{prompt: "Download?", yes: true, width: 40}
+	got := m.View()
+	// Just ensure it doesn't panic and returns content
+	if got == "" {
+		t.Error("View with width set should still return content")
+	}
+}

cmd/tui/selector.go

@@ -0,0 +1,654 @@
+package tui
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+)
+
+var (
+	selectorTitleStyle = lipgloss.NewStyle().
+				Bold(true)
+
+	selectorItemStyle = lipgloss.NewStyle().
+				PaddingLeft(4)
+
+	selectorSelectedItemStyle = lipgloss.NewStyle().
+					PaddingLeft(2).
+					Bold(true).
+					Background(lipgloss.AdaptiveColor{Light: "254", Dark: "236"})
+
+	selectorDescStyle = lipgloss.NewStyle().
+				Foreground(lipgloss.AdaptiveColor{Light: "242", Dark: "246"})
+
+	selectorDescLineStyle = selectorDescStyle.
+				PaddingLeft(6)
+
+	selectorFilterStyle = lipgloss.NewStyle().
+				Foreground(lipgloss.AdaptiveColor{Light: "242", Dark: "246"}).
+				Italic(true)
+
+	selectorInputStyle = lipgloss.NewStyle().
+				Foreground(lipgloss.AdaptiveColor{Light: "235", Dark: "252"})
+
+	selectorCheckboxStyle = lipgloss.NewStyle().
+				Foreground(lipgloss.AdaptiveColor{Light: "242", Dark: "246"})
+
+	selectorCheckboxCheckedStyle = lipgloss.NewStyle().
+					Bold(true)
+
+	selectorDefaultTagStyle = lipgloss.NewStyle().
+				Foreground(lipgloss.AdaptiveColor{Light: "242", Dark: "246"}).
+				Italic(true)
+
+	selectorHelpStyle = lipgloss.NewStyle().
+				Foreground(lipgloss.AdaptiveColor{Light: "244", Dark: "244"})
+
+	selectorMoreStyle = lipgloss.NewStyle().
+				PaddingLeft(6).
+				Foreground(lipgloss.AdaptiveColor{Light: "242", Dark: "246"}).
+				Italic(true)
+
+	sectionHeaderStyle = lipgloss.NewStyle().
+				PaddingLeft(2).
+				Bold(true).
+				Foreground(lipgloss.AdaptiveColor{Light: "240", Dark: "249"})
+)
+
+const maxSelectorItems = 10
+
+// ErrCancelled is returned when the user cancels the selection.
+var ErrCancelled = errors.New("cancelled")
+
+type SelectItem struct {
+	Name        string
+	Description string
+	Recommended bool
+}
+
+// selectorModel is the bubbletea model for single selection.
+type selectorModel struct {
+	title        string
+	items        []SelectItem
+	filter       string
+	cursor       int
+	scrollOffset int
+	selected     string
+	cancelled    bool
+	helpText     string
+	width        int
+}
+
+func (m selectorModel) filteredItems() []SelectItem {
+	if m.filter == "" {
+		return m.items
+	}
+	filterLower := strings.ToLower(m.filter)
+	var result []SelectItem
+	for _, item := range m.items {
+		if strings.Contains(strings.ToLower(item.Name), filterLower) {
+			result = append(result, item)
+		}
+	}
+	return result
+}
+
+func (m selectorModel) Init() tea.Cmd {
+	return nil
+}
+
+// otherStart returns the index of the first non-recommended item in the filtered list.
+// When filtering, all items scroll together so this returns 0.
+func (m selectorModel) otherStart() int {
+	if m.filter != "" {
+		return 0
+	}
+	filtered := m.filteredItems()
+	for i, item := range filtered {
+		if !item.Recommended {
+			return i
+		}
+	}
+	return len(filtered)
+}
+
+// updateNavigation handles navigation keys (up/down/pgup/pgdown/filter/backspace).
+// It does NOT handle Enter, Esc, or CtrlC. This is used by both the standalone
+// selector and the TUI modal (which intercepts Enter/Esc for its own logic).
+func (m *selectorModel) updateNavigation(msg tea.KeyMsg) {
+	filtered := m.filteredItems()
+	otherStart := m.otherStart()
+
+	switch msg.Type {
+	case tea.KeyUp:
+		if m.cursor > 0 {
+			m.cursor--
+			m.updateScroll(otherStart)
+		}
+
+	case tea.KeyDown:
+		if m.cursor < len(filtered)-1 {
+			m.cursor++
+			m.updateScroll(otherStart)
+		}
+
+	case tea.KeyPgUp:
+		m.cursor -= maxSelectorItems
+		if m.cursor < 0 {
+			m.cursor = 0
+		}
+		m.updateScroll(otherStart)
+
+	case tea.KeyPgDown:
+		m.cursor += maxSelectorItems
+		if m.cursor >= len(filtered) {
+			m.cursor = len(filtered) - 1
+		}
+		m.updateScroll(otherStart)
+
+	case tea.KeyBackspace:
+		if len(m.filter) > 0 {
+			m.filter = m.filter[:len(m.filter)-1]
+			m.cursor = 0
+			m.scrollOffset = 0
+		}
+
+	case tea.KeyRunes:
+		m.filter += string(msg.Runes)
+		m.cursor = 0
+		m.scrollOffset = 0
+	}
+}
+
+// updateScroll adjusts scrollOffset based on cursor position.
+// When not filtering, scrollOffset is relative to the "More" (non-recommended) section.
+// When filtering, it's relative to the full filtered list.
+func (m *selectorModel) updateScroll(otherStart int) {
+	if m.filter != "" {
+		if m.cursor < m.scrollOffset {
+			m.scrollOffset = m.cursor
+		}
+		if m.cursor >= m.scrollOffset+maxSelectorItems {
+			m.scrollOffset = m.cursor - maxSelectorItems + 1
+		}
+		return
+	}
+
+	// Cursor is in recommended section — reset "More" scroll to top
+	if m.cursor < otherStart {
+		m.scrollOffset = 0
+		return
+	}
+
+	// Cursor is in "More" section — scroll relative to others
+	posInOthers := m.cursor - otherStart
+	maxOthers := maxSelectorItems - otherStart
+	if maxOthers < 3 {
+		maxOthers = 3
+	}
+	if posInOthers < m.scrollOffset {
+		m.scrollOffset = posInOthers
+	}
+	if posInOthers >= m.scrollOffset+maxOthers {
+		m.scrollOffset = posInOthers - maxOthers + 1
+	}
+}
+
+func (m selectorModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
+	switch msg := msg.(type) {
+	case tea.WindowSizeMsg:
+		wasSet := m.width > 0
+		m.width = msg.Width
+		if wasSet {
+			return m, tea.EnterAltScreen
+		}
+		return m, nil
+
+	case tea.KeyMsg:
+		switch msg.Type {
+		case tea.KeyCtrlC, tea.KeyEsc:
+			m.cancelled = true
+			return m, tea.Quit
+
+		case tea.KeyEnter:
+			filtered := m.filteredItems()
+			if len(filtered) > 0 && m.cursor < len(filtered) {
+				m.selected = filtered[m.cursor].Name
+			}
+			return m, tea.Quit
+
+		default:
+			m.updateNavigation(msg)
+		}
+	}
+
+	return m, nil
+}
+
+func (m selectorModel) renderItem(s *strings.Builder, item SelectItem, idx int) {
+	if idx == m.cursor {
+		s.WriteString(selectorSelectedItemStyle.Render("▸ " + item.Name))
+	} else {
+		s.WriteString(selectorItemStyle.Render(item.Name))
+	}
+	s.WriteString("\n")
+	if item.Description != "" {
+		s.WriteString(selectorDescLineStyle.Render(item.Description))
+		s.WriteString("\n")
+	}
+}
+
+// renderContent renders the selector content (title, items, help text) without
+// checking the cancelled/selected state. This is used by both View() (standalone mode)
+// and by the TUI modal which embeds a selectorModel.
+func (m selectorModel) renderContent() string {
+	var s strings.Builder
+
+	s.WriteString(selectorTitleStyle.Render(m.title))
+	s.WriteString(" ")
+	if m.filter == "" {
+		s.WriteString(selectorFilterStyle.Render("Type to filter..."))
+	} else {
+		s.WriteString(selectorInputStyle.Render(m.filter))
+	}
+	s.WriteString("\n\n")
+
+	filtered := m.filteredItems()
+
+	if len(filtered) == 0 {
+		s.WriteString(selectorItemStyle.Render(selectorDescStyle.Render("(no matches)")))
+		s.WriteString("\n")
+	} else if m.filter != "" {
+		s.WriteString(sectionHeaderStyle.Render("Top Results"))
+		s.WriteString("\n")
+
+		displayCount := min(len(filtered), maxSelectorItems)
+		for i := range displayCount {
+			idx := m.scrollOffset + i
+			if idx >= len(filtered) {
+				break
+			}
+			m.renderItem(&s, filtered[idx], idx)
+		}
+
+		if remaining := len(filtered) - m.scrollOffset - displayCount; remaining > 0 {
+			s.WriteString(selectorMoreStyle.Render(fmt.Sprintf("... and %d more", remaining)))
+			s.WriteString("\n")
+		}
+	} else {
+		// Split into pinned recommended and scrollable others
+		var recItems, otherItems []int
+		for i, item := range filtered {
+			if item.Recommended {
+				recItems = append(recItems, i)
+			} else {
+				otherItems = append(otherItems, i)
+			}
+		}
+
+		// Always render all recommended items (pinned)
+		if len(recItems) > 0 {
+			s.WriteString(sectionHeaderStyle.Render("Recommended"))
+			s.WriteString("\n")
+			for _, idx := range recItems {
+				m.renderItem(&s, filtered[idx], idx)
+			}
+		}
+
+		if len(otherItems) > 0 {
+			s.WriteString("\n")
+			s.WriteString(sectionHeaderStyle.Render("More"))
+			s.WriteString("\n")
+
+			maxOthers := maxSelectorItems - len(recItems)
+			if maxOthers < 3 {
+				maxOthers = 3
+			}
+			displayCount := min(len(otherItems), maxOthers)
+
+			for i := range displayCount {
+				idx := m.scrollOffset + i
+				if idx >= len(otherItems) {
+					break
+				}
+				m.renderItem(&s, filtered[otherItems[idx]], otherItems[idx])
+			}
+
+			if remaining := len(otherItems) - m.scrollOffset - displayCount; remaining > 0 {
+				s.WriteString(selectorMoreStyle.Render(fmt.Sprintf("... and %d more", remaining)))
+				s.WriteString("\n")
+			}
+		}
+	}
+
+	s.WriteString("\n")
+	help := "↑/↓ navigate • enter select • esc cancel"
+	if m.helpText != "" {
+		help = m.helpText
+	}
+	s.WriteString(selectorHelpStyle.Render(help))
+
+	return s.String()
+}
+
+func (m selectorModel) View() string {
+	if m.cancelled || m.selected != "" {
+		return ""
+	}
+
+	s := m.renderContent()
+	if m.width > 0 {
+		return lipgloss.NewStyle().MaxWidth(m.width).Render(s)
+	}
+	return s
+}
+
+func SelectSingle(title string, items []SelectItem) (string, error) {
+	if len(items) == 0 {
+		return "", fmt.Errorf("no items to select from")
+	}
+
+	m := selectorModel{
+		title: title,
+		items: items,
+	}
+
+	p := tea.NewProgram(m)
+	finalModel, err := p.Run()
+	if err != nil {
+		return "", fmt.Errorf("error running selector: %w", err)
+	}
+
+	fm := finalModel.(selectorModel)
+	if fm.cancelled {
+		return "", ErrCancelled
+	}
+
+	return fm.selected, nil
+}
+
+// multiSelectorModel is the bubbletea model for multi selection.
+type multiSelectorModel struct {
+	title        string
+	items        []SelectItem
+	itemIndex    map[string]int
+	filter       string
+	cursor       int
+	scrollOffset int
+	checked      map[int]bool
+	checkOrder   []int
+	cancelled    bool
+	confirmed    bool
+	width        int
+}
+
+func newMultiSelectorModel(title string, items []SelectItem, preChecked []string) multiSelectorModel {
+	m := multiSelectorModel{
+		title:     title,
+		items:     items,
+		itemIndex: make(map[string]int, len(items)),
+		checked:   make(map[int]bool),
+	}
+
+	for i, item := range items {
+		m.itemIndex[item.Name] = i
+	}
+
+	for _, name := range preChecked {
+		if idx, ok := m.itemIndex[name]; ok {
+			m.checked[idx] = true
+			m.checkOrder = append(m.checkOrder, idx)
+		}
+	}
+
+	return m
+}
+
+func (m multiSelectorModel) filteredItems() []SelectItem {
+	if m.filter == "" {
+		return m.items
+	}
+	filterLower := strings.ToLower(m.filter)
+	var result []SelectItem
+	for _, item := range m.items {
+		if strings.Contains(strings.ToLower(item.Name), filterLower) {
+			result = append(result, item)
+		}
+	}
+	return result
+}
+
+func (m *multiSelectorModel) toggleItem() {
+	filtered := m.filteredItems()
+	if len(filtered) == 0 || m.cursor >= len(filtered) {
+		return
+	}
+
+	item := filtered[m.cursor]
+	origIdx := m.itemIndex[item.Name]
+
+	if m.checked[origIdx] {
+		delete(m.checked, origIdx)
+		for i, idx := range m.checkOrder {
+			if idx == origIdx {
+				m.checkOrder = append(m.checkOrder[:i], m.checkOrder[i+1:]...)
+				break
+			}
+		}
+	} else {
+		m.checked[origIdx] = true
+		m.checkOrder = append(m.checkOrder, origIdx)
+	}
+}
+
+func (m multiSelectorModel) selectedCount() int {
+	return len(m.checkOrder)
+}
+
+func (m multiSelectorModel) Init() tea.Cmd {
+	return nil
+}
+
+func (m multiSelectorModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
+	switch msg := msg.(type) {
+	case tea.WindowSizeMsg:
+		wasSet := m.width > 0
+		m.width = msg.Width
+		if wasSet {
+			return m, tea.EnterAltScreen
+		}
+		return m, nil
+
+	case tea.KeyMsg:
+		filtered := m.filteredItems()
+
+		switch msg.Type {
+		case tea.KeyCtrlC, tea.KeyEsc:
+			m.cancelled = true
+			return m, tea.Quit
+
+		case tea.KeyEnter:
+			if len(m.checkOrder) > 0 {
+				m.confirmed = true
+				return m, tea.Quit
+			}
+
+		case tea.KeySpace:
+			m.toggleItem()
+
+		case tea.KeyUp:
+			if m.cursor > 0 {
+				m.cursor--
+				if m.cursor < m.scrollOffset {
+					m.scrollOffset = m.cursor
+				}
+			}
+
+		case tea.KeyDown:
+			if m.cursor < len(filtered)-1 {
+				m.cursor++
+				if m.cursor >= m.scrollOffset+maxSelectorItems {
+					m.scrollOffset = m.cursor - maxSelectorItems + 1
+				}
+			}
+
+		case tea.KeyPgUp:
+			m.cursor -= maxSelectorItems
+			if m.cursor < 0 {
+				m.cursor = 0
+			}
+			m.scrollOffset -= maxSelectorItems
+			if m.scrollOffset < 0 {
+				m.scrollOffset = 0
+			}
+
+		case tea.KeyPgDown:
+			m.cursor += maxSelectorItems
+			if m.cursor >= len(filtered) {
+				m.cursor = len(filtered) - 1
+			}
+			if m.cursor >= m.scrollOffset+maxSelectorItems {
+				m.scrollOffset = m.cursor - maxSelectorItems + 1
+			}
+
+		case tea.KeyBackspace:
+			if len(m.filter) > 0 {
+				m.filter = m.filter[:len(m.filter)-1]
+				m.cursor = 0
+				m.scrollOffset = 0
+			}
+
+		case tea.KeyRunes:
+			m.filter += string(msg.Runes)
+			m.cursor = 0
+			m.scrollOffset = 0
+		}
+	}
+
+	return m, nil
+}
+
+func (m multiSelectorModel) View() string {
+	if m.cancelled || m.confirmed {
+		return ""
+	}
+
+	var s strings.Builder
+
+	s.WriteString(selectorTitleStyle.Render(m.title))
+	s.WriteString(" ")
+	if m.filter == "" {
+		s.WriteString(selectorFilterStyle.Render("Type to filter..."))
+	} else {
+		s.WriteString(selectorInputStyle.Render(m.filter))
+	}
+	s.WriteString("\n\n")
+
+	filtered := m.filteredItems()
+
+	if len(filtered) == 0 {
+		s.WriteString(selectorItemStyle.Render(selectorDescStyle.Render("(no matches)")))
+		s.WriteString("\n")
+	} else {
+		displayCount := min(len(filtered), maxSelectorItems)
+		shownRecHeader := false
+		prevWasRec := false
+
+		for i := range displayCount {
+			idx := m.scrollOffset + i
+			if idx >= len(filtered) {
+				break
+			}
+			item := filtered[idx]
+			origIdx := m.itemIndex[item.Name]
+
+			if m.filter == "" {
+				if item.Recommended && !shownRecHeader {
+					s.WriteString(sectionHeaderStyle.Render("Recommended"))
+					s.WriteString("\n")
+					shownRecHeader = true
+				} else if !item.Recommended && prevWasRec {
+					s.WriteString("\n")
+				}
+				prevWasRec = item.Recommended
+			}
+
+			var checkbox string
+			if m.checked[origIdx] {
+				checkbox = selectorCheckboxCheckedStyle.Render("[x]")
+			} else {
+				checkbox = selectorCheckboxStyle.Render("[ ]")
+			}
+
+			var line string
+			if idx == m.cursor {
+				line = selectorSelectedItemStyle.Render("▸ ") + checkbox + " " + selectorSelectedItemStyle.Render(item.Name)
+			} else {
+				line = "  " + checkbox + " " + item.Name
+			}
+
+			if len(m.checkOrder) > 0 && m.checkOrder[0] == origIdx {
+				line += " " + selectorDefaultTagStyle.Render("(default)")
+			}
+
+			s.WriteString(line)
+			s.WriteString("\n")
+		}
+
+		if remaining := len(filtered) - m.scrollOffset - displayCount; remaining > 0 {
+			s.WriteString(selectorMoreStyle.Render(fmt.Sprintf("... and %d more", remaining)))
+			s.WriteString("\n")
+		}
+	}
+
+	s.WriteString("\n")
+
+	count := m.selectedCount()
+	if count == 0 {
+		s.WriteString(selectorDescStyle.Render("  Select at least one model."))
+	} else {
+		s.WriteString(selectorDescStyle.Render(fmt.Sprintf("  %d selected - press enter to continue", count)))
+	}
+	s.WriteString("\n\n")
+
+	s.WriteString(selectorHelpStyle.Render("↑/↓ navigate • space toggle • enter confirm • esc cancel"))
+
+	result := s.String()
+	if m.width > 0 {
+		return lipgloss.NewStyle().MaxWidth(m.width).Render(result)
+	}
+	return result
+}
+
+func SelectMultiple(title string, items []SelectItem, preChecked []string) ([]string, error) {
+	if len(items) == 0 {
+		return nil, fmt.Errorf("no items to select from")
+	}
+
+	m := newMultiSelectorModel(title, items, preChecked)
+
+	p := tea.NewProgram(m)
+	finalModel, err := p.Run()
+	if err != nil {
+		return nil, fmt.Errorf("error running selector: %w", err)
+	}
+
+	fm := finalModel.(multiSelectorModel)
+	if fm.cancelled {
+		return nil, ErrCancelled
+	}
+
+	if !fm.confirmed {
+		return nil, ErrCancelled
+	}
+
+	var result []string
+	for _, idx := range fm.checkOrder {
+		result = append(result, fm.items[idx].Name)
+	}
+
+	return result, nil
+}

cmd/tui/selector_test.go

@@ -0,0 +1,410 @@
+package tui
+
+import (
+	"strings"
+	"testing"
+
+	tea "github.com/charmbracelet/bubbletea"
+)
+
+func items(names ...string) []SelectItem {
+	var out []SelectItem
+	for _, n := range names {
+		out = append(out, SelectItem{Name: n})
+	}
+	return out
+}
+
+func recItems(names ...string) []SelectItem {
+	var out []SelectItem
+	for _, n := range names {
+		out = append(out, SelectItem{Name: n, Recommended: true})
+	}
+	return out
+}
+
+func mixedItems() []SelectItem {
+	return []SelectItem{
+		{Name: "rec-a", Recommended: true},
+		{Name: "rec-b", Recommended: true},
+		{Name: "other-1"},
+		{Name: "other-2"},
+		{Name: "other-3"},
+		{Name: "other-4"},
+		{Name: "other-5"},
+		{Name: "other-6"},
+		{Name: "other-7"},
+		{Name: "other-8"},
+		{Name: "other-9"},
+		{Name: "other-10"},
+	}
+}
+
+func TestFilteredItems(t *testing.T) {
+	tests := []struct {
+		name   string
+		items  []SelectItem
+		filter string
+		want   []string
+	}{
+		{
+			name:   "no filter returns all",
+			items:  items("alpha", "beta", "gamma"),
+			filter: "",
+			want:   []string{"alpha", "beta", "gamma"},
+		},
+		{
+			name:   "filter matches substring",
+			items:  items("llama3.2", "qwen3:8b", "llama2"),
+			filter: "llama",
+			want:   []string{"llama3.2", "llama2"},
+		},
+		{
+			name:   "filter is case insensitive",
+			items:  items("Qwen3:8b", "llama3.2"),
+			filter: "QWEN",
+			want:   []string{"Qwen3:8b"},
+		},
+		{
+			name:   "no matches",
+			items:  items("alpha", "beta"),
+			filter: "zzz",
+			want:   nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			m := selectorModel{items: tt.items, filter: tt.filter}
+			got := m.filteredItems()
+			var gotNames []string
+			for _, item := range got {
+				gotNames = append(gotNames, item.Name)
+			}
+			if len(gotNames) != len(tt.want) {
+				t.Fatalf("got %v, want %v", gotNames, tt.want)
+			}
+			for i := range tt.want {
+				if gotNames[i] != tt.want[i] {
+					t.Errorf("index %d: got %q, want %q", i, gotNames[i], tt.want[i])
+				}
+			}
+		})
+	}
+}
+
+func TestOtherStart(t *testing.T) {
+	tests := []struct {
+		name   string
+		items  []SelectItem
+		filter string
+		want   int
+	}{
+		{
+			name:  "all recommended",
+			items: recItems("a", "b", "c"),
+			want:  3,
+		},
+		{
+			name:  "none recommended",
+			items: items("a", "b"),
+			want:  0,
+		},
+		{
+			name: "mixed",
+			items: []SelectItem{
+				{Name: "rec-a", Recommended: true},
+				{Name: "rec-b", Recommended: true},
+				{Name: "other-1"},
+				{Name: "other-2"},
+			},
+			want: 2,
+		},
+		{
+			name:  "empty",
+			items: nil,
+			want:  0,
+		},
+		{
+			name: "filtering returns 0",
+			items: []SelectItem{
+				{Name: "rec-a", Recommended: true},
+				{Name: "other-1"},
+			},
+			filter: "rec",
+			want:   0,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			m := selectorModel{items: tt.items, filter: tt.filter}
+			if got := m.otherStart(); got != tt.want {
+				t.Errorf("otherStart() = %d, want %d", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestUpdateScroll(t *testing.T) {
+	tests := []struct {
+		name       string
+		cursor     int
+		offset     int
+		otherStart int
+		filter     string
+		wantOffset int
+	}{
+		{
+			name:       "cursor in recommended resets scroll",
+			cursor:     1,
+			offset:     5,
+			otherStart: 3,
+			wantOffset: 0,
+		},
+		{
+			name:       "cursor at start of others",
+			cursor:     2,
+			offset:     0,
+			otherStart: 2,
+			wantOffset: 0,
+		},
+		{
+			name:       "cursor scrolls down in others",
+			cursor:     12,
+			offset:     0,
+			otherStart: 2,
+			wantOffset: 3, // posInOthers=10, maxOthers=8, 10-8+1=3
+		},
+		{
+			name:       "cursor scrolls up in others",
+			cursor:     4,
+			offset:     5,
+			otherStart: 2,
+			wantOffset: 2, // posInOthers=2 < offset=5
+		},
+		{
+			name:       "filter mode standard scroll down",
+			cursor:     12,
+			offset:     0,
+			filter:     "x",
+			otherStart: 0,
+			wantOffset: 3, // 12 - 10 + 1 = 3
+		},
+		{
+			name:       "filter mode standard scroll up",
+			cursor:     2,
+			offset:     5,
+			filter:     "x",
+			otherStart: 0,
+			wantOffset: 2,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			m := selectorModel{
+				cursor:       tt.cursor,
+				scrollOffset: tt.offset,
+				filter:       tt.filter,
+			}
+			m.updateScroll(tt.otherStart)
+			if m.scrollOffset != tt.wantOffset {
+				t.Errorf("scrollOffset = %d, want %d", m.scrollOffset, tt.wantOffset)
+			}
+		})
+	}
+}
+
+func TestRenderContent_SectionHeaders(t *testing.T) {
+	m := selectorModel{
+		title: "Pick:",
+		items: []SelectItem{
+			{Name: "rec-a", Recommended: true},
+			{Name: "other-1"},
+		},
+	}
+	content := m.renderContent()
+
+	if !strings.Contains(content, "Recommended") {
+		t.Error("should contain 'Recommended' header")
+	}
+	if !strings.Contains(content, "More") {
+		t.Error("should contain 'More' header")
+	}
+}
+
+func TestRenderContent_FilteredHeader(t *testing.T) {
+	m := selectorModel{
+		title:  "Pick:",
+		items:  items("alpha", "beta", "alphabet"),
+		filter: "alpha",
+	}
+	content := m.renderContent()
+
+	if !strings.Contains(content, "Top Results") {
+		t.Error("filtered view should contain 'Top Results' header")
+	}
+	if strings.Contains(content, "Recommended") {
+		t.Error("filtered view should not contain 'Recommended' header")
+	}
+}
+
+func TestRenderContent_NoMatches(t *testing.T) {
+	m := selectorModel{
+		title:  "Pick:",
+		items:  items("alpha"),
+		filter: "zzz",
+	}
+	content := m.renderContent()
+
+	if !strings.Contains(content, "(no matches)") {
+		t.Error("should show '(no matches)' when filter has no results")
+	}
+}
+
+func TestRenderContent_SelectedItemIndicator(t *testing.T) {
+	m := selectorModel{
+		title:  "Pick:",
+		items:  items("alpha", "beta"),
+		cursor: 0,
+	}
+	content := m.renderContent()
+
+	if !strings.Contains(content, "▸") {
+		t.Error("selected item should have ▸ indicator")
+	}
+}
+
+func TestRenderContent_Description(t *testing.T) {
+	m := selectorModel{
+		title: "Pick:",
+		items: []SelectItem{
+			{Name: "alpha", Description: "the first letter"},
+		},
+	}
+	content := m.renderContent()
+
+	if !strings.Contains(content, "the first letter") {
+		t.Error("should render item description")
+	}
+}
+
+func TestRenderContent_PinnedRecommended(t *testing.T) {
+	m := selectorModel{
+		title: "Pick:",
+		items: mixedItems(),
+		// cursor deep in "More" section
+		cursor:       8,
+		scrollOffset: 3,
+	}
+	content := m.renderContent()
+
+	// Recommended items should always be visible (pinned)
+	if !strings.Contains(content, "rec-a") {
+		t.Error("recommended items should always be rendered (pinned)")
+	}
+	if !strings.Contains(content, "rec-b") {
+		t.Error("recommended items should always be rendered (pinned)")
+	}
+}
+
+func TestRenderContent_MoreOverflowIndicator(t *testing.T) {
+	m := selectorModel{
+		title: "Pick:",
+		items: mixedItems(), // 2 rec + 10 other = 12 total, maxSelectorItems=10
+	}
+	content := m.renderContent()
+
+	if !strings.Contains(content, "... and") {
+		t.Error("should show overflow indicator when more items than visible")
+	}
+}
+
+func TestUpdateNavigation_CursorBounds(t *testing.T) {
+	m := selectorModel{
+		items:  items("a", "b", "c"),
+		cursor: 0,
+	}
+
+	// Up at top stays at 0
+	m.updateNavigation(keyMsg(KeyUp))
+	if m.cursor != 0 {
+		t.Errorf("cursor should stay at 0 when pressing up at top, got %d", m.cursor)
+	}
+
+	// Down moves to 1
+	m.updateNavigation(keyMsg(KeyDown))
+	if m.cursor != 1 {
+		t.Errorf("cursor should be 1 after down, got %d", m.cursor)
+	}
+
+	// Down to end
+	m.updateNavigation(keyMsg(KeyDown))
+	m.updateNavigation(keyMsg(KeyDown))
+	if m.cursor != 2 {
+		t.Errorf("cursor should be 2 at bottom, got %d", m.cursor)
+	}
+}
+
+func TestUpdateNavigation_FilterResetsState(t *testing.T) {
+	m := selectorModel{
+		items:        items("alpha", "beta"),
+		cursor:       1,
+		scrollOffset: 5,
+	}
+
+	m.updateNavigation(runeMsg('x'))
+	if m.filter != "x" {
+		t.Errorf("filter should be 'x', got %q", m.filter)
+	}
+	if m.cursor != 0 {
+		t.Errorf("cursor should reset to 0 on filter, got %d", m.cursor)
+	}
+	if m.scrollOffset != 0 {
+		t.Errorf("scrollOffset should reset to 0 on filter, got %d", m.scrollOffset)
+	}
+}
+
+func TestUpdateNavigation_Backspace(t *testing.T) {
+	m := selectorModel{
+		items:  items("alpha"),
+		filter: "abc",
+		cursor: 1,
+	}
+
+	m.updateNavigation(keyMsg(KeyBackspace))
+	if m.filter != "ab" {
+		t.Errorf("filter should be 'ab' after backspace, got %q", m.filter)
+	}
+	if m.cursor != 0 {
+		t.Errorf("cursor should reset to 0 on backspace, got %d", m.cursor)
+	}
+}
+
+// Key message helpers for testing
+
+type keyType = int
+
+const (
+	KeyUp        keyType = iota
+	KeyDown      keyType = iota
+	KeyBackspace keyType = iota
+)
+
+func keyMsg(k keyType) tea.KeyMsg {
+	switch k {
+	case KeyUp:
+		return tea.KeyMsg{Type: tea.KeyUp}
+	case KeyDown:
+		return tea.KeyMsg{Type: tea.KeyDown}
+	case KeyBackspace:
+		return tea.KeyMsg{Type: tea.KeyBackspace}
+	default:
+		return tea.KeyMsg{}
+	}
+}
+
+func runeMsg(r rune) tea.KeyMsg {
+	return tea.KeyMsg{Type: tea.KeyRunes, Runes: []rune{r}}
+}

cmd/tui/signin.go

@@ -0,0 +1,128 @@
+package tui
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+	"github.com/ollama/ollama/cmd/config"
+)
+
+type signInModel struct {
+	modelName string
+	signInURL string
+	spinner   int
+	width     int
+	userName  string
+	cancelled bool
+}
+
+func (m signInModel) Init() tea.Cmd {
+	return tea.Tick(200*time.Millisecond, func(t time.Time) tea.Msg {
+		return signInTickMsg{}
+	})
+}
+
+func (m signInModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
+	switch msg := msg.(type) {
+	case tea.WindowSizeMsg:
+		wasSet := m.width > 0
+		m.width = msg.Width
+		if wasSet {
+			return m, tea.EnterAltScreen
+		}
+		return m, nil
+
+	case tea.KeyMsg:
+		switch msg.Type {
+		case tea.KeyCtrlC, tea.KeyEsc:
+			m.cancelled = true
+			return m, tea.Quit
+		}
+
+	case signInTickMsg:
+		m.spinner++
+		if m.spinner%5 == 0 {
+			return m, tea.Batch(
+				tea.Tick(200*time.Millisecond, func(t time.Time) tea.Msg {
+					return signInTickMsg{}
+				}),
+				checkSignIn,
+			)
+		}
+		return m, tea.Tick(200*time.Millisecond, func(t time.Time) tea.Msg {
+			return signInTickMsg{}
+		})
+
+	case signInCheckMsg:
+		if msg.signedIn {
+			m.userName = msg.userName
+			return m, tea.Quit
+		}
+	}
+
+	return m, nil
+}
+
+func (m signInModel) View() string {
+	if m.userName != "" {
+		return ""
+	}
+	return renderSignIn(m.modelName, m.signInURL, m.spinner, m.width)
+}
+
+func renderSignIn(modelName, signInURL string, spinner, width int) string {
+	spinnerFrames := []string{"⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"}
+	frame := spinnerFrames[spinner%len(spinnerFrames)]
+
+	urlColor := lipgloss.NewStyle().
+		Foreground(lipgloss.Color("117"))
+	urlWrap := lipgloss.NewStyle().PaddingLeft(2)
+	if width > 4 {
+		urlWrap = urlWrap.Width(width - 4)
+	}
+
+	var s strings.Builder
+
+	fmt.Fprintf(&s, "To use %s, please sign in.\n\n", selectorSelectedItemStyle.Render(modelName))
+
+	// Wrap in OSC 8 hyperlink so the entire URL is clickable even when wrapped.
+	// Padding is outside the hyperlink so spaces don't get underlined.
+	link := fmt.Sprintf("\033]8;;%s\033\\%s\033]8;;\033\\", signInURL, urlColor.Render(signInURL))
+	s.WriteString("Navigate to:\n")
+	s.WriteString(urlWrap.Render(link))
+	s.WriteString("\n\n")
+
+	s.WriteString(lipgloss.NewStyle().Foreground(lipgloss.AdaptiveColor{Light: "242", Dark: "246"}).Render(
+		frame + " Waiting for sign in to complete..."))
+	s.WriteString("\n\n")
+
+	s.WriteString(selectorHelpStyle.Render("esc cancel"))
+
+	return lipgloss.NewStyle().PaddingLeft(2).Render(s.String())
+}
+
+// RunSignIn shows a bubbletea sign-in dialog and polls until the user signs in or cancels.
+func RunSignIn(modelName, signInURL string) (string, error) {
+	config.OpenBrowser(signInURL)
+
+	m := signInModel{
+		modelName: modelName,
+		signInURL: signInURL,
+	}
+
+	p := tea.NewProgram(m)
+	finalModel, err := p.Run()
+	if err != nil {
+		return "", fmt.Errorf("error running sign-in: %w", err)
+	}
+
+	fm := finalModel.(signInModel)
+	if fm.cancelled {
+		return "", ErrCancelled
+	}
+
+	return fm.userName, nil
+}

cmd/tui/signin_test.go

@@ -0,0 +1,175 @@
+package tui
+
+import (
+	"strings"
+	"testing"
+
+	tea "github.com/charmbracelet/bubbletea"
+)
+
+func TestRenderSignIn_ContainsModelName(t *testing.T) {
+	got := renderSignIn("glm-4.7:cloud", "https://example.com/signin", 0, 80)
+	if !strings.Contains(got, "glm-4.7:cloud") {
+		t.Error("should contain model name")
+	}
+	if !strings.Contains(got, "please sign in") {
+		t.Error("should contain sign-in prompt")
+	}
+}
+
+func TestRenderSignIn_ContainsURL(t *testing.T) {
+	url := "https://ollama.com/connect?key=abc123"
+	got := renderSignIn("test:cloud", url, 0, 120)
+	if !strings.Contains(got, url) {
+		t.Errorf("should contain URL %q", url)
+	}
+}
+
+func TestRenderSignIn_OSC8Hyperlink(t *testing.T) {
+	url := "https://ollama.com/connect?key=abc123"
+	got := renderSignIn("test:cloud", url, 0, 120)
+
+	// Should contain OSC 8 open sequence with the URL
+	osc8Open := "\033]8;;" + url + "\033\\"
+	if !strings.Contains(got, osc8Open) {
+		t.Error("should contain OSC 8 open sequence with URL")
+	}
+
+	// Should contain OSC 8 close sequence
+	osc8Close := "\033]8;;\033\\"
+	if !strings.Contains(got, osc8Close) {
+		t.Error("should contain OSC 8 close sequence")
+	}
+}
+
+func TestRenderSignIn_ContainsSpinner(t *testing.T) {
+	got := renderSignIn("test:cloud", "https://example.com", 0, 80)
+	if !strings.Contains(got, "Waiting for sign in to complete") {
+		t.Error("should contain waiting message")
+	}
+	if !strings.Contains(got, "⠋") {
+		t.Error("should contain first spinner frame at spinner=0")
+	}
+}
+
+func TestRenderSignIn_SpinnerAdvances(t *testing.T) {
+	got0 := renderSignIn("test:cloud", "https://example.com", 0, 80)
+	got1 := renderSignIn("test:cloud", "https://example.com", 1, 80)
+	if got0 == got1 {
+		t.Error("different spinner values should produce different output")
+	}
+}
+
+func TestRenderSignIn_ContainsEscHelp(t *testing.T) {
+	got := renderSignIn("test:cloud", "https://example.com", 0, 80)
+	if !strings.Contains(got, "esc cancel") {
+		t.Error("should contain esc cancel help text")
+	}
+}
+
+func TestSignInModel_EscCancels(t *testing.T) {
+	m := signInModel{
+		modelName: "test:cloud",
+		signInURL: "https://example.com",
+	}
+
+	updated, cmd := m.Update(tea.KeyMsg{Type: tea.KeyEsc})
+	fm := updated.(signInModel)
+	if !fm.cancelled {
+		t.Error("esc should set cancelled=true")
+	}
+	if cmd == nil {
+		t.Error("esc should return tea.Quit")
+	}
+}
+
+func TestSignInModel_CtrlCCancels(t *testing.T) {
+	m := signInModel{
+		modelName: "test:cloud",
+		signInURL: "https://example.com",
+	}
+
+	updated, cmd := m.Update(tea.KeyMsg{Type: tea.KeyCtrlC})
+	fm := updated.(signInModel)
+	if !fm.cancelled {
+		t.Error("ctrl+c should set cancelled=true")
+	}
+	if cmd == nil {
+		t.Error("ctrl+c should return tea.Quit")
+	}
+}
+
+func TestSignInModel_SignedInQuitsClean(t *testing.T) {
+	m := signInModel{
+		modelName: "test:cloud",
+		signInURL: "https://example.com",
+	}
+
+	updated, cmd := m.Update(signInCheckMsg{signedIn: true, userName: "alice"})
+	fm := updated.(signInModel)
+	if fm.userName != "alice" {
+		t.Errorf("expected userName 'alice', got %q", fm.userName)
+	}
+	if cmd == nil {
+		t.Error("successful sign-in should return tea.Quit")
+	}
+}
+
+func TestSignInModel_SignedInViewClears(t *testing.T) {
+	m := signInModel{
+		modelName: "test:cloud",
+		signInURL: "https://example.com",
+		userName:  "alice",
+	}
+
+	got := m.View()
+	if got != "" {
+		t.Errorf("View should return empty string after sign-in, got %q", got)
+	}
+}
+
+func TestSignInModel_NotSignedInContinues(t *testing.T) {
+	m := signInModel{
+		modelName: "test:cloud",
+		signInURL: "https://example.com",
+	}
+
+	updated, _ := m.Update(signInCheckMsg{signedIn: false})
+	fm := updated.(signInModel)
+	if fm.userName != "" {
+		t.Error("should not set userName when not signed in")
+	}
+	if fm.cancelled {
+		t.Error("should not cancel when check returns not signed in")
+	}
+}
+
+func TestSignInModel_WindowSizeUpdatesWidth(t *testing.T) {
+	m := signInModel{
+		modelName: "test:cloud",
+		signInURL: "https://example.com",
+	}
+
+	updated, _ := m.Update(tea.WindowSizeMsg{Width: 120, Height: 40})
+	fm := updated.(signInModel)
+	if fm.width != 120 {
+		t.Errorf("expected width 120, got %d", fm.width)
+	}
+}
+
+func TestSignInModel_TickAdvancesSpinner(t *testing.T) {
+	m := signInModel{
+		modelName: "test:cloud",
+		signInURL: "https://example.com",
+		spinner:   0,
+	}
+
+	updated, cmd := m.Update(signInTickMsg{})
+	fm := updated.(signInModel)
+	if fm.spinner != 1 {
+		t.Errorf("expected spinner=1, got %d", fm.spinner)
+	}
+	if cmd == nil {
+		t.Error("tick should return a command")
+	}
+}

cmd/tui/tui.go

@@ -0,0 +1,603 @@
+package tui
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/cmd/config"
+	"github.com/ollama/ollama/version"
+)
+
+var (
+	versionStyle = lipgloss.NewStyle().
+			Foreground(lipgloss.AdaptiveColor{Light: "243", Dark: "250"})
+
+	menuItemStyle = lipgloss.NewStyle().
+			PaddingLeft(2)
+
+	menuSelectedItemStyle = lipgloss.NewStyle().
+				Bold(true).
+				Background(lipgloss.AdaptiveColor{Light: "254", Dark: "236"})
+
+	menuDescStyle = selectorDescStyle.
+			PaddingLeft(4)
+
+	greyedStyle = menuItemStyle.
+			Foreground(lipgloss.AdaptiveColor{Light: "242", Dark: "246"})
+
+	greyedSelectedStyle = menuSelectedItemStyle.
+				Foreground(lipgloss.AdaptiveColor{Light: "242", Dark: "246"})
+
+	modelStyle = lipgloss.NewStyle().
+			Foreground(lipgloss.AdaptiveColor{Light: "243", Dark: "250"})
+
+	notInstalledStyle = lipgloss.NewStyle().
+				Foreground(lipgloss.AdaptiveColor{Light: "242", Dark: "246"}).
+				Italic(true)
+)
+
+type menuItem struct {
+	title       string
+	description string
+	integration string // integration name for loading model config, empty if not an integration
+	isRunModel  bool
+	isOthers    bool
+}
+
+var mainMenuItems = []menuItem{
+	{
+		title:       "Run a model",
+		description: "Start an interactive chat with a model",
+		isRunModel:  true,
+	},
+	{
+		title:       "Launch Claude Code",
+		description: "Agentic coding across large codebases",
+		integration: "claude",
+	},
+	{
+		title:       "Launch Codex",
+		description: "OpenAI's open-source coding agent",
+		integration: "codex",
+	},
+	{
+		title:       "Launch OpenClaw",
+		description: "Personal AI with 100+ skills",
+		integration: "openclaw",
+	},
+}
+
+var othersMenuItem = menuItem{
+	title:       "More...",
+	description: "Show additional integrations",
+	isOthers:    true,
+}
+
+// getOtherIntegrations dynamically builds the "Others" list from the integration
+// registry, excluding any integrations already present in the pinned mainMenuItems.
+func getOtherIntegrations() []menuItem {
+	pinned := map[string]bool{
+		"run": true, // not an integration but in the pinned list
+	}
+	for _, item := range mainMenuItems {
+		if item.integration != "" {
+			pinned[item.integration] = true
+		}
+	}
+
+	var others []menuItem
+	for _, info := range config.ListIntegrationInfos() {
+		if pinned[info.Name] {
+			continue
+		}
+		desc := info.Description
+		if desc == "" {
+			desc = "Open " + info.DisplayName + " integration"
+		}
+		others = append(others, menuItem{
+			title:       "Launch " + info.DisplayName,
+			description: desc,
+			integration: info.Name,
+		})
+	}
+	return others
+}
+
+type model struct {
+	items           []menuItem
+	cursor          int
+	quitting        bool
+	selected        bool
+	changeModel     bool
+	showOthers      bool
+	availableModels map[string]bool
+	err             error
+
+	showingModal  bool
+	modalSelector selectorModel
+	modalItems    []SelectItem
+
+	showingSignIn   bool
+	signInURL       string
+	signInModel     string
+	signInSpinner   int
+	signInFromModal bool   // true if sign-in was triggered from modal (not main menu)
+
+	width     int    // terminal width from WindowSizeMsg
+	statusMsg string // temporary status message shown near help text
+}
+
+type signInTickMsg struct{}
+
+type signInCheckMsg struct {
+	signedIn bool
+	userName string
+}
+
+type clearStatusMsg struct{}
+
+func (m *model) modelExists(name string) bool {
+	if m.availableModels == nil || name == "" {
+		return false
+	}
+	if m.availableModels[name] {
+		return true
+	}
+	// Check for prefix match (e.g., "llama2" matches "llama2:latest")
+	for modelName := range m.availableModels {
+		if strings.HasPrefix(modelName, name+":") {
+			return true
+		}
+	}
+	return false
+}
+
+func (m *model) buildModalItems() []SelectItem {
+	modelItems, _ := config.GetModelItems(context.Background())
+	var items []SelectItem
+	for _, item := range modelItems {
+		items = append(items, SelectItem{Name: item.Name, Description: item.Description, Recommended: item.Recommended})
+	}
+	return items
+}
+
+func (m *model) openModelModal() {
+	m.modalItems = m.buildModalItems()
+	m.modalSelector = selectorModel{
+		title:    "Select model:",
+		items:    m.modalItems,
+		helpText: "↑/↓ navigate • enter select • ← back",
+	}
+	m.showingModal = true
+}
+
+func isCloudModel(name string) bool {
+	return strings.HasSuffix(name, ":cloud")
+}
+
+// checkCloudSignIn checks if a cloud model needs sign-in.
+// Returns a command to start sign-in if needed, or nil if already signed in.
+func (m *model) checkCloudSignIn(modelName string, fromModal bool) tea.Cmd {
+	if modelName == "" || !isCloudModel(modelName) {
+		return nil
+	}
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return nil
+	}
+	user, err := client.Whoami(context.Background())
+	if err == nil && user != nil && user.Name != "" {
+		return nil
+	}
+	var aErr api.AuthorizationError
+	if errors.As(err, &aErr) && aErr.SigninURL != "" {
+		return m.startSignIn(modelName, aErr.SigninURL, fromModal)
+	}
+	return nil
+}
+
+// startSignIn initiates the sign-in flow for a cloud model.
+// fromModal indicates if this was triggered from the model picker modal.
+func (m *model) startSignIn(modelName, signInURL string, fromModal bool) tea.Cmd {
+	m.showingModal = false
+	m.showingSignIn = true
+	m.signInURL = signInURL
+	m.signInModel = modelName
+	m.signInSpinner = 0
+	m.signInFromModal = fromModal
+
+	config.OpenBrowser(signInURL)
+
+	return tea.Tick(200*time.Millisecond, func(t time.Time) tea.Msg {
+		return signInTickMsg{}
+	})
+}
+
+func checkSignIn() tea.Msg {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return signInCheckMsg{signedIn: false}
+	}
+	user, err := client.Whoami(context.Background())
+	if err == nil && user != nil && user.Name != "" {
+		return signInCheckMsg{signedIn: true, userName: user.Name}
+	}
+	return signInCheckMsg{signedIn: false}
+}
+
+func (m *model) loadAvailableModels() {
+	m.availableModels = make(map[string]bool)
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return
+	}
+	models, err := client.List(context.Background())
+	if err != nil {
+		return
+	}
+	for _, mdl := range models.Models {
+		m.availableModels[mdl.Name] = true
+	}
+}
+
+func (m *model) buildItems() {
+	others := getOtherIntegrations()
+	m.items = make([]menuItem, 0, len(mainMenuItems)+1+len(others))
+	m.items = append(m.items, mainMenuItems...)
+
+	if m.showOthers {
+		m.items = append(m.items, others...)
+	} else {
+		m.items = append(m.items, othersMenuItem)
+	}
+}
+
+func isOthersIntegration(name string) bool {
+	for _, item := range getOtherIntegrations() {
+		if item.integration == name {
+			return true
+		}
+	}
+	return false
+}
+
+func initialModel() model {
+	m := model{
+		cursor: 0,
+	}
+	m.loadAvailableModels()
+
+	lastSelection := config.LastSelection()
+	if isOthersIntegration(lastSelection) {
+		m.showOthers = true
+	}
+
+	m.buildItems()
+
+	if lastSelection != "" {
+		for i, item := range m.items {
+			if lastSelection == "run" && item.isRunModel {
+				m.cursor = i
+				break
+			} else if item.integration == lastSelection {
+				m.cursor = i
+				break
+			}
+		}
+	}
+
+	return m
+}
+
+func (m model) Init() tea.Cmd {
+	return nil
+}
+
+func (m model) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
+	if wmsg, ok := msg.(tea.WindowSizeMsg); ok {
+		wasSet := m.width > 0
+		m.width = wmsg.Width
+		if wasSet {
+			return m, tea.EnterAltScreen
+		}
+		return m, nil
+	}
+
+	if _, ok := msg.(clearStatusMsg); ok {
+		m.statusMsg = ""
+		return m, nil
+	}
+
+	if m.showingSignIn {
+		switch msg := msg.(type) {
+		case tea.KeyMsg:
+			switch msg.Type {
+			case tea.KeyCtrlC, tea.KeyEsc:
+				m.showingSignIn = false
+				if m.signInFromModal {
+					m.showingModal = true
+				}
+				return m, nil
+			}
+
+		case signInTickMsg:
+			m.signInSpinner++
+			// Check sign-in status every 5th tick (~1 second)
+			if m.signInSpinner%5 == 0 {
+				return m, tea.Batch(
+					tea.Tick(200*time.Millisecond, func(t time.Time) tea.Msg {
+						return signInTickMsg{}
+					}),
+					checkSignIn,
+				)
+			}
+			return m, tea.Tick(200*time.Millisecond, func(t time.Time) tea.Msg {
+				return signInTickMsg{}
+			})
+
+		case signInCheckMsg:
+			if msg.signedIn {
+				if m.signInFromModal {
+					m.modalSelector.selected = m.signInModel
+					m.changeModel = true
+				} else {
+					m.selected = true
+				}
+				m.quitting = true
+				return m, tea.Quit
+			}
+		}
+		return m, nil
+	}
+
+	if m.showingModal {
+		switch msg := msg.(type) {
+		case tea.KeyMsg:
+			switch msg.Type {
+			case tea.KeyCtrlC, tea.KeyEsc, tea.KeyLeft:
+				m.showingModal = false
+				return m, nil
+
+			case tea.KeyEnter:
+				filtered := m.modalSelector.filteredItems()
+				if len(filtered) > 0 && m.modalSelector.cursor < len(filtered) {
+					m.modalSelector.selected = filtered[m.modalSelector.cursor].Name
+				}
+				if m.modalSelector.selected != "" {
+					if cmd := m.checkCloudSignIn(m.modalSelector.selected, true); cmd != nil {
+						return m, cmd
+					}
+					m.changeModel = true
+					m.quitting = true
+					return m, tea.Quit
+				}
+				return m, nil
+
+			default:
+				// Delegate navigation (up/down/pgup/pgdown/filter/backspace) to selectorModel
+				m.modalSelector.updateNavigation(msg)
+			}
+		}
+		return m, nil
+	}
+
+	switch msg := msg.(type) {
+	case tea.KeyMsg:
+		switch msg.String() {
+		case "ctrl+c", "q", "esc":
+			m.quitting = true
+			return m, tea.Quit
+
+		case "up", "k":
+			if m.cursor > 0 {
+				m.cursor--
+			}
+			// Auto-collapse "Others" when cursor moves back into pinned items
+			if m.showOthers && m.cursor < len(mainMenuItems) {
+				m.showOthers = false
+				m.buildItems()
+			}
+
+		case "down", "j":
+			if m.cursor < len(m.items)-1 {
+				m.cursor++
+			}
+			// Auto-expand "Others..." when cursor lands on it
+			if m.cursor < len(m.items) && m.items[m.cursor].isOthers && !m.showOthers {
+				m.showOthers = true
+				m.buildItems()
+				// cursor now points at the first "other" integration
+			}
+
+		case "enter", " ":
+			item := m.items[m.cursor]
+
+			if item.integration != "" && !config.IsIntegrationInstalled(item.integration) {
+				return m, nil
+			}
+
+			var configuredModel string
+			if item.isRunModel {
+				configuredModel = config.LastModel()
+			} else if item.integration != "" {
+				configuredModel = config.IntegrationModel(item.integration)
+			}
+			if cmd := m.checkCloudSignIn(configuredModel, false); cmd != nil {
+				return m, cmd
+			}
+
+			m.selected = true
+			m.quitting = true
+			return m, tea.Quit
+
+		case "right", "l":
+			item := m.items[m.cursor]
+			if item.integration != "" || item.isRunModel {
+				if item.integration != "" && !config.IsIntegrationInstalled(item.integration) {
+					return m, nil
+				}
+				m.openModelModal()
+			}
+		}
+	}
+
+	return m, nil
+}
+
+func (m model) View() string {
+	if m.quitting {
+		return ""
+	}
+
+	if m.showingSignIn {
+		return m.renderSignInDialog()
+	}
+
+	if m.showingModal {
+		return m.renderModal()
+	}
+
+	s := selectorTitleStyle.Render("Ollama "+versionStyle.Render(version.Version)) + "\n\n"
+
+	for i, item := range m.items {
+		cursor := ""
+		style := menuItemStyle
+		isInstalled := true
+
+		if item.integration != "" {
+			isInstalled = config.IsIntegrationInstalled(item.integration)
+		}
+
+		if m.cursor == i {
+			cursor = "▸ "
+			if isInstalled {
+				style = menuSelectedItemStyle
+			} else {
+				style = greyedSelectedStyle
+			}
+		} else if !isInstalled && item.integration != "" {
+			style = greyedStyle
+		}
+
+		title := item.title
+		var modelSuffix string
+		if item.integration != "" {
+			if !isInstalled {
+				title += " " + notInstalledStyle.Render("(not installed)")
+			} else if m.cursor == i {
+				if mdl := config.IntegrationModel(item.integration); mdl != "" && m.modelExists(mdl) {
+					modelSuffix = " " + modelStyle.Render("("+mdl+")")
+				}
+			}
+		} else if item.isRunModel && m.cursor == i {
+			if mdl := config.LastModel(); mdl != "" && m.modelExists(mdl) {
+				modelSuffix = " " + modelStyle.Render("("+mdl+")")
+			}
+		}
+
+		s += style.Render(cursor+title) + modelSuffix + "\n"
+
+		desc := item.description
+		if !isInstalled && item.integration != "" && m.cursor == i {
+			if hint := config.IntegrationInstallHint(item.integration); hint != "" {
+				desc = hint
+			} else {
+				desc = "not installed"
+			}
+		}
+		s += menuDescStyle.Render(desc) + "\n\n"
+	}
+
+	if m.statusMsg != "" {
+		s += "\n" + lipgloss.NewStyle().Foreground(lipgloss.AdaptiveColor{Light: "124", Dark: "210"}).Render(m.statusMsg) + "\n"
+	}
+
+	s += "\n" + selectorHelpStyle.Render("↑/↓ navigate • enter launch • → change model • esc quit")
+
+	if m.width > 0 {
+		return lipgloss.NewStyle().MaxWidth(m.width).Render(s)
+	}
+	return s
+}
+
+func (m model) renderModal() string {
+	modalStyle := lipgloss.NewStyle().
+		PaddingBottom(1).
+		PaddingRight(2)
+
+	s := modalStyle.Render(m.modalSelector.renderContent())
+	if m.width > 0 {
+		return lipgloss.NewStyle().MaxWidth(m.width).Render(s)
+	}
+	return s
+}
+
+func (m model) renderSignInDialog() string {
+	return renderSignIn(m.signInModel, m.signInURL, m.signInSpinner, m.width)
+}
+
+type Selection int
+
+const (
+	SelectionNone Selection = iota
+	SelectionRunModel
+	SelectionChangeRunModel
+	SelectionIntegration       // Generic integration selection
+	SelectionChangeIntegration // Generic change model for integration
+)
+
+type Result struct {
+	Selection   Selection
+	Integration string // integration name if applicable
+	Model       string // model name if selected from modal
+}
+
+func Run() (Result, error) {
+	m := initialModel()
+	p := tea.NewProgram(m)
+
+	finalModel, err := p.Run()
+	if err != nil {
+		return Result{Selection: SelectionNone}, fmt.Errorf("error running TUI: %w", err)
+	}
+
+	fm := finalModel.(model)
+	if fm.err != nil {
+		return Result{Selection: SelectionNone}, fm.err
+	}
+
+	if !fm.selected && !fm.changeModel {
+		return Result{Selection: SelectionNone}, nil
+	}
+
+	item := fm.items[fm.cursor]
+
+	if fm.changeModel {
+		if item.isRunModel {
+			return Result{
+				Selection: SelectionChangeRunModel,
+				Model:     fm.modalSelector.selected,
+			}, nil
+		}
+		return Result{
+			Selection:   SelectionChangeIntegration,
+			Integration: item.integration,
+			Model:       fm.modalSelector.selected,
+		}, nil
+	}
+
+	if item.isRunModel {
+		return Result{Selection: SelectionRunModel}, nil
+	}
+
+	return Result{
+		Selection:   SelectionIntegration,
+		Integration: item.integration,
+	}, nil
+}

convert/convert.go

@@ -311,6 +311,14 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 		conv = &deepseekocr{}
 	case "DeepseekV3ForCausalLM":
 		conv = &deepseek2Model{}
+	case "Glm4MoeLiteForCausalLM":
+		conv = &glm4MoeLiteModel{}
+	case "GlmOcrForConditionalGeneration":
+		conv = &glmOcrModel{}
+	case "Lfm2ForCausalLM":
+		conv = &lfm2Model{}
+	case "Qwen3NextForCausalLM":
+		conv = &qwen3NextModel{}
 	default:
 		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}

convert/convert_glm4moelite.go

@@ -0,0 +1,264 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"log/slog"
+	"regexp"
+	"strconv"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"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["glm4moelite.attention.key_length_mla"] = p.KVLoraRank + p.QKRopeHeadDim
+	kv["glm4moelite.attention.value_length_mla"] = p.KVLoraRank
+
+	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",
+	}
+}
+
+// repackKVB extracts K or V from the combined KV_B tensor for MLA absorption.
+// K output row-major: [n_head, kv_lora_rank, qk_nope] -> GGML ne[]={qk_nope, kv_lora_rank, n_head}
+// V output row-major: [n_head, v_head, kv_lora_rank] -> GGML ne[]={kv_lora_rank, v_head, n_head}
+func (p *glm4MoeLiteModel) repackKVB(extractK bool, kvFirst bool, numHeads int) Repacker {
+	qkNope := int(p.QKNopeHeadDim)
+	vHeadDim := int(p.VHeadDim)
+	kvLoraRank := int(p.KVLoraRank)
+	kvPerHead := qkNope + vHeadDim
+
+	return func(_ string, data []float32, shape []uint64) ([]float32, error) {
+		dims := make([]int, len(shape))
+		for i := range shape {
+			dims[i] = int(shape[i])
+		}
+
+		var tt tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+		var err error
+
+		// Normalize to [n_head * (qk_nope + v_head), kv_lora_rank] layout
+		if kvFirst {
+			tt, err = tensor.Transpose(tt, 1, 0)
+			if err != nil {
+				return nil, err
+			}
+			tt = tensor.Materialize(tt)
+		}
+
+		// Reshape to [n_head, qk_nope + v_head, kv_lora_rank]
+		if err := tt.Reshape(numHeads, kvPerHead, kvLoraRank); err != nil {
+			return nil, err
+		}
+
+		if extractK {
+			// Slice K: [n_head, qk_nope, kv_lora_rank]
+			tt, err = tt.Slice(nil, tensor.S(0, qkNope), nil)
+			if err != nil {
+				return nil, err
+			}
+			tt = tensor.Materialize(tt)
+			// Transpose to [n_head, kv_lora_rank, qk_nope]
+			tt, err = tensor.Transpose(tt, 0, 2, 1)
+			if err != nil {
+				return nil, err
+			}
+			tt = tensor.Materialize(tt)
+		} else {
+			// Slice V: [n_head, v_head, kv_lora_rank] - already correct layout
+			tt, err = tt.Slice(nil, tensor.S(qkNope, kvPerHead), nil)
+			if err != nil {
+				return nil, err
+			}
+			tt = tensor.Materialize(tt)
+		}
+
+		if err := tt.Reshape(tt.Shape().TotalSize()); err != nil {
+			return nil, err
+		}
+		return native.VectorF32(tt.(*tensor.Dense))
+	}
+}
+
+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
+		}
+
+		// Split attn_kv_b into separate attn_k_b and attn_v_b for MLA absorption
+		if strings.HasSuffix(t.Name(), ".attn_kv_b.weight") {
+			qkNope := int(p.QKNopeHeadDim)
+			vHeadDim := int(p.VHeadDim)
+			kvLoraRank := int(p.KVLoraRank)
+			kvPerHead := qkNope + vHeadDim
+			numHeads := int(p.NumAttentionHeads)
+			kvFirst := true
+			if len(t.Shape()) == 2 {
+				switch {
+				case int(t.Shape()[0]) == kvLoraRank:
+					if kvPerHead > 0 && int(t.Shape()[1])%kvPerHead == 0 {
+						numHeads = int(t.Shape()[1]) / kvPerHead
+					}
+					kvFirst = true
+				case int(t.Shape()[1]) == kvLoraRank:
+					if kvPerHead > 0 && int(t.Shape()[0])%kvPerHead == 0 {
+						numHeads = int(t.Shape()[0]) / kvPerHead
+					}
+					kvFirst = false
+				default:
+					slog.Warn("glm4moelite: unexpected attn_kv_b layout", "name", t.Name(), "shape", t.Shape())
+				}
+			}
+
+			kTensor := t.Clone()
+			kTensor.SetRepacker(p.repackKVB(true, kvFirst, numHeads))
+			out = append(out, &ggml.Tensor{
+				Name:     strings.Replace(t.Name(), "attn_kv_b", "attn_k_b", 1),
+				Kind:     t.Kind(),
+				Shape:    []uint64{uint64(numHeads), uint64(kvLoraRank), uint64(qkNope)},
+				WriterTo: kTensor,
+			})
+
+			vTensor := t.Clone()
+			vTensor.SetRepacker(p.repackKVB(false, kvFirst, numHeads))
+			out = append(out, &ggml.Tensor{
+				Name:     strings.Replace(t.Name(), "attn_kv_b", "attn_v_b", 1),
+				Kind:     t.Kind(),
+				Shape:    []uint64{uint64(numHeads), uint64(vHeadDim), uint64(kvLoraRank)},
+				WriterTo: vTensor,
+			})
+			continue
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+	return out
+}

convert/convert_glmocr.go

@@ -0,0 +1,455 @@
+package convert
+
+import (
+	"cmp"
+	"encoding/json"
+	"io/fs"
+	"log/slog"
+	"regexp"
+	"strconv"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+)
+
+// normalToNeoXRepacker creates a repacker that permutes Q/K weights from interleaved (LLaMA)
+// to NeoX ordering for compatibility with GGML's M-RoPE kernel.
+//
+// For weights: reshape [out, in] -> [n_heads, head_dim, in], permute rotary dims, reshape back
+// For biases: reshape [out] -> [n_heads, head_dim], permute rotary dims, reshape back
+func normalToNeoXRepacker(nHeads, headDim int, partialRotaryFactor float32) func(string, []float32, []uint64) ([]float32, error) {
+	return func(_ string, data []float32, shape []uint64) ([]float32, error) {
+		rotaryDim := int(float32(headDim) * partialRotaryFactor)
+		if rotaryDim%2 != 0 {
+			rotaryDim = (rotaryDim / 2) * 2 // Round down to even
+		}
+
+		// Handle 1D (bias) or 2D (weight) tensors
+		is1D := len(shape) == 1
+		var inFeatures int
+		if is1D {
+			inFeatures = 1
+		} else {
+			inFeatures = int(shape[1])
+		}
+		outFeatures := int(shape[0])
+		nEffectiveHeads := outFeatures / headDim
+
+		if nEffectiveHeads != nHeads {
+			slog.Warn("normalToNeoX: unexpected head count", "effective", nEffectiveHeads, "expected", nHeads)
+		}
+
+		// Reshape to [n_heads, head_dim, in_features]
+		reshaped := make([]float32, len(data))
+		copy(reshaped, data)
+
+		// Permute the rotary dimensions: even indices first, then odd
+		// For each head, reorder [0,1,2,3,4,5...] to [0,2,4...,1,3,5...]
+		result := make([]float32, len(data))
+		halfRotary := rotaryDim / 2
+
+		for h := range nEffectiveHeads {
+			for f := range inFeatures {
+				for i := range halfRotary {
+					// Even dim (0, 2, 4, ...) -> position i
+					srcIdx := h*headDim*inFeatures + (2*i)*inFeatures + f
+					dstIdx := h*headDim*inFeatures + i*inFeatures + f
+					result[dstIdx] = reshaped[srcIdx]
+
+					// Odd dim (1, 3, 5, ...) -> position halfRotary + i
+					srcIdx = h*headDim*inFeatures + (2*i+1)*inFeatures + f
+					dstIdx = h*headDim*inFeatures + (halfRotary+i)*inFeatures + f
+					result[dstIdx] = reshaped[srcIdx]
+				}
+
+				// Non-rotary part: copy as-is
+				for i := rotaryDim; i < headDim; i++ {
+					srcIdx := h*headDim*inFeatures + i*inFeatures + f
+					result[srcIdx] = reshaped[srcIdx]
+				}
+			}
+		}
+
+		return result, nil
+	}
+}
+
+type glmOcrModel struct {
+	ModelParameters
+
+	TextConfig struct {
+		HiddenSize          uint32  `json:"hidden_size"`
+		IntermediateSize    uint32  `json:"intermediate_size"`
+		NumHiddenLayers     uint32  `json:"num_hidden_layers"`
+		NumAttentionHeads   uint32  `json:"num_attention_heads"`
+		NumKeyValueHeads    uint32  `json:"num_key_value_heads"`
+		HeadDim             uint32  `json:"head_dim"`
+		MaxPositionEmbed    uint32  `json:"max_position_embeddings"`
+		RMSNormEps          float32 `json:"rms_norm_eps"`
+		PartialRotaryFactor float32 `json:"partial_rotary_factor"`
+		RopeParameters      struct {
+			RopeType            string  `json:"rope_type"`
+			MRopeSection        []int32 `json:"mrope_section"`
+			RopeTheta           float32 `json:"rope_theta"`
+			PartialRotaryFactor float32 `json:"partial_rotary_factor"`
+		} `json:"rope_parameters"`
+	} `json:"text_config"`
+
+	VisionConfig struct {
+		HiddenSize        uint32  `json:"hidden_size"`
+		IntermediateSize  uint32  `json:"intermediate_size"`
+		Depth             uint32  `json:"depth"`
+		NumHeads          uint32  `json:"num_heads"`
+		ImageSize         uint32  `json:"image_size"`
+		PatchSize         uint32  `json:"patch_size"`
+		OutHiddenSize     uint32  `json:"out_hidden_size"`
+		RMSNormEps        float32 `json:"rms_norm_eps"`
+		SpatialMergeSize  uint32  `json:"spatial_merge_size"`
+		TemporalPatchSize uint32  `json:"temporal_patch_size"`
+	} `json:"vision_config"`
+
+	ImageStartTokenID uint32 `json:"image_start_token_id"`
+	ImageEndTokenID   uint32 `json:"image_end_token_id"`
+	VideoStartTokenID uint32 `json:"video_start_token_id"`
+	VideoEndTokenID   uint32 `json:"video_end_token_id"`
+	ImageTokenID      uint32 `json:"image_token_id"`
+	VideoTokenID      uint32 `json:"video_token_id"`
+
+	// Preprocessor config (preprocessor_config.json)
+	Preprocessor struct {
+		Size struct {
+			ShortestEdge uint32 `json:"shortest_edge"`
+			LongestEdge  uint32 `json:"longest_edge"`
+		} `json:"size"`
+		PatchSize         uint32    `json:"patch_size"`
+		TemporalPatchSize uint32    `json:"temporal_patch_size"`
+		MergeSize         uint32    `json:"merge_size"`
+		ImageMean         []float32 `json:"image_mean"`
+		ImageStd          []float32 `json:"image_std"`
+	} `json:"-"`
+}
+
+var _ ModelConverter = (*glmOcrModel)(nil)
+
+func (m *glmOcrModel) parseMore(fsys fs.FS) error {
+	bts, err := fs.ReadFile(fsys, "preprocessor_config.json")
+	if err != nil {
+		return err
+	}
+
+	return json.Unmarshal(bts, &m.Preprocessor)
+}
+
+func (m *glmOcrModel) KV(t *Tokenizer) KV {
+	kv := m.ModelParameters.KV(t)
+	kv["general.architecture"] = "glmocr"
+
+	// Text model parameters
+	kv["glmocr.block_count"] = cmp.Or(m.TextConfig.NumHiddenLayers, 16)
+	kv["glmocr.embedding_length"] = cmp.Or(m.TextConfig.HiddenSize, 1536)
+	kv["glmocr.attention.head_count"] = cmp.Or(m.TextConfig.NumAttentionHeads, 16)
+	kv["glmocr.attention.head_count_kv"] = cmp.Or(m.TextConfig.NumKeyValueHeads, 8)
+	headDim := cmp.Or(m.TextConfig.HeadDim, m.TextConfig.HiddenSize/m.TextConfig.NumAttentionHeads)
+	kv["glmocr.attention.key_length"] = headDim
+	kv["glmocr.attention.value_length"] = headDim
+	kv["glmocr.feed_forward_length"] = cmp.Or(m.TextConfig.IntermediateSize, 4608)
+	kv["glmocr.attention.layer_norm_rms_epsilon"] = cmp.Or(m.TextConfig.RMSNormEps, 1e-5)
+	kv["glmocr.context_length"] = cmp.Or(m.TextConfig.MaxPositionEmbed, 131072)
+	kv["glmocr.rope.freq_base"] = cmp.Or(m.TextConfig.RopeParameters.RopeTheta, float32(10000))
+	kv["glmocr.rope.partial_rotary_factor"] = cmp.Or(m.TextConfig.RopeParameters.PartialRotaryFactor, m.TextConfig.PartialRotaryFactor, float32(1.0))
+	if len(m.TextConfig.RopeParameters.MRopeSection) > 0 {
+		kv["glmocr.rope.mrope_section"] = m.TextConfig.RopeParameters.MRopeSection
+	}
+
+	// Vision model parameters
+	kv["glmocr.vision.block_count"] = cmp.Or(m.VisionConfig.Depth, 24)
+	kv["glmocr.vision.embedding_length"] = cmp.Or(m.VisionConfig.HiddenSize, 1024)
+	kv["glmocr.vision.attention.head_count"] = cmp.Or(m.VisionConfig.NumHeads, 16)
+	kv["glmocr.vision.image_size"] = cmp.Or(m.VisionConfig.ImageSize, 336)
+	kv["glmocr.vision.patch_size"] = cmp.Or(m.VisionConfig.PatchSize, m.Preprocessor.PatchSize, 14)
+	kv["glmocr.vision.spatial_merge_size"] = cmp.Or(m.VisionConfig.SpatialMergeSize, m.Preprocessor.MergeSize, 2)
+	kv["glmocr.vision.temporal_patch_size"] = cmp.Or(m.VisionConfig.TemporalPatchSize, m.Preprocessor.TemporalPatchSize, 2)
+	kv["glmocr.vision.out_hidden_size"] = cmp.Or(m.VisionConfig.OutHiddenSize, 1536)
+	kv["glmocr.vision.intermediate_size"] = cmp.Or(m.VisionConfig.IntermediateSize, 4096)
+	kv["glmocr.vision.attention.layer_norm_rms_epsilon"] = cmp.Or(m.VisionConfig.RMSNormEps, 1e-5)
+
+	// Preprocessor-derived image settings (min/max pixels and normalization)
+	// Note: fs.Config.keyValue() auto-prepends architecture prefix, so use full key
+	if m.Preprocessor.Size.ShortestEdge > 0 {
+		kv["glmocr.vision.min_pixels"] = m.Preprocessor.Size.ShortestEdge
+	}
+	if m.Preprocessor.Size.LongestEdge > 0 {
+		kv["glmocr.vision.max_pixels"] = m.Preprocessor.Size.LongestEdge
+	}
+	if len(m.Preprocessor.ImageMean) == 3 {
+		kv["glmocr.vision.image_mean"] = m.Preprocessor.ImageMean
+	}
+	if len(m.Preprocessor.ImageStd) == 3 {
+		kv["glmocr.vision.image_std"] = m.Preprocessor.ImageStd
+	}
+
+	// Special tokens
+	kv["glmocr.image_token_id"] = m.ImageTokenID
+	kv["glmocr.image_start_token_id"] = m.ImageStartTokenID
+	kv["glmocr.image_end_token_id"] = m.ImageEndTokenID
+	kv["glmocr.video_token_id"] = m.VideoTokenID
+	kv["glmocr.video_start_token_id"] = m.VideoStartTokenID
+	kv["glmocr.video_end_token_id"] = m.VideoEndTokenID
+
+	return kv
+}
+
+func (m *glmOcrModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+
+	// Skip layers >= num_hidden_layers (Multi-Token Prediction layers not needed for basic inference)
+	numLayers := int(cmp.Or(m.TextConfig.NumHiddenLayers, 16))
+	skipLayer := func(name string) bool {
+		// Tensor names are already replaced to "blk.N.xxx" format
+		re := regexp.MustCompile(`^blk\.(\d+)`)
+		matches := re.FindStringSubmatch(name)
+		if matches == nil {
+			return false
+		}
+		blkNum, err := strconv.Atoi(matches[1])
+		if err != nil {
+			return false
+		}
+		return blkNum >= numLayers
+	}
+
+	for _, t := range ts {
+		name := t.Name()
+
+		// Skip next-n prediction layers (layers >= num_hidden_layers)
+		if skipLayer(name) {
+			continue
+		}
+
+		// Split ffn_gate_up into separate gate and up projections
+		if strings.Contains(name, "ffn_gate_up") {
+			for t := range splitDim(t, 0,
+				split{Replacer: strings.NewReplacer("ffn_gate_up", "ffn_gate")},
+				split{Replacer: strings.NewReplacer("ffn_gate_up", "ffn_up")},
+			) {
+				out = append(out, t)
+			}
+			continue
+		}
+
+		if strings.HasSuffix(name, "patch_embd.weight") {
+			shape := t.Shape()
+			if len(shape) == 5 && shape[2] == 2 {
+				newShape := []uint64{shape[0], shape[1], shape[3], shape[4]}
+
+				t0 := t.Clone()
+				t0.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
+					dims := make([]int, len(shape))
+					for i := range shape {
+						dims[i] = int(shape[i])
+					}
+					var tt tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+					tt, err := tt.Slice(nil, nil, tensor.S(0, 1), nil, nil)
+					if err != nil {
+						return nil, err
+					}
+					tt = tensor.Materialize(tt)
+					newDims := []int{int(shape[0]), int(shape[1]), int(shape[3]), int(shape[4])}
+					if err := tt.Reshape(newDims...); err != nil {
+						return nil, err
+					}
+					if err := tt.Reshape(tt.Shape().TotalSize()); err != nil {
+						return nil, err
+					}
+					return native.VectorF32(tt.(*tensor.Dense))
+				})
+				out = append(out, &ggml.Tensor{
+					Name:     strings.Replace(name, "patch_embd.weight", "patch_embd_0.weight", 1),
+					Kind:     t.Kind(),
+					Shape:    newShape,
+					WriterTo: t0,
+				})
+
+				t1 := t.Clone()
+				t1.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
+					dims := make([]int, len(shape))
+					for i := range shape {
+						dims[i] = int(shape[i])
+					}
+					var tt tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+					tt, err := tt.Slice(nil, nil, tensor.S(1, 2), nil, nil)
+					if err != nil {
+						return nil, err
+					}
+					tt = tensor.Materialize(tt)
+					newDims := []int{int(shape[0]), int(shape[1]), int(shape[3]), int(shape[4])}
+					if err := tt.Reshape(newDims...); err != nil {
+						return nil, err
+					}
+					if err := tt.Reshape(tt.Shape().TotalSize()); err != nil {
+						return nil, err
+					}
+					return native.VectorF32(tt.(*tensor.Dense))
+				})
+				out = append(out, &ggml.Tensor{
+					Name:     strings.Replace(name, "patch_embd.weight", "patch_embd_1.weight", 1),
+					Kind:     t.Kind(),
+					Shape:    newShape,
+					WriterTo: t1,
+				})
+
+				continue
+			}
+
+			if len(shape) == 4 {
+				out = append(out, &ggml.Tensor{
+					Name:     strings.Replace(name, "patch_embd.weight", "patch_embd_0.weight", 1),
+					Kind:     t.Kind(),
+					Shape:    t.Shape(),
+					WriterTo: t,
+				})
+				continue
+			}
+
+			slog.Warn("glmocr: patch_embed weight has unexpected shape - not splitting", "shape", shape)
+			// Fall through to default handling
+		}
+
+		// Handle pre-split patch embedding weights
+		// Pattern 1: v.patch_embd.0.weight, v.patch_embd.1.weight -> patch_embd_0.weight, patch_embd_1.weight
+		// Pattern 2: v.patch_embd.weight.0, v.patch_embd.weight.1 -> patch_embd_0.weight, patch_embd_1.weight
+		if strings.Contains(name, "patch_embd.0.") {
+			out = append(out, &ggml.Tensor{
+				Name:     strings.Replace(name, "patch_embd.0.", "patch_embd_0.", 1),
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: t,
+			})
+			continue
+		}
+		if strings.Contains(name, "patch_embd.1.") {
+			out = append(out, &ggml.Tensor{
+				Name:     strings.Replace(name, "patch_embd.1.", "patch_embd_1.", 1),
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: t,
+			})
+			continue
+		}
+		// Handle .weight.0 and .weight.1 suffix patterns
+		if strings.HasSuffix(name, "patch_embd.weight.0") {
+			out = append(out, &ggml.Tensor{
+				Name:     strings.Replace(name, "patch_embd.weight.0", "patch_embd_0.weight", 1),
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: t,
+			})
+			continue
+		}
+		if strings.HasSuffix(name, "patch_embd.weight.1") {
+			out = append(out, &ggml.Tensor{
+				Name:     strings.Replace(name, "patch_embd.weight.1", "patch_embd_1.weight", 1),
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: t,
+			})
+			continue
+		}
+
+		// Permute Q/K weights for M-RoPE compatibility (interleaved -> NeoX ordering)
+		// GGML's M-RoPE kernel uses NeoX-style rotation, but GLM-OCR uses interleaved (LLaMA-style)
+		// We permute at conversion time so the weights work correctly with GGML's kernel
+		// This aligns Q/K rotary dimensions with GGML's NeoX-style rotation
+		if len(m.TextConfig.RopeParameters.MRopeSection) > 0 &&
+			strings.Contains(name, "blk.") && (strings.Contains(name, "attn_q.") || strings.Contains(name, "attn_k.")) {
+			// Get config values for permutation
+			nHeads := int(cmp.Or(m.TextConfig.NumAttentionHeads, 16))
+			nKVHeads := int(cmp.Or(m.TextConfig.NumKeyValueHeads, 8))
+			hiddenSize := int(cmp.Or(m.TextConfig.HiddenSize, 1536))
+			headDim := int(cmp.Or(m.TextConfig.HeadDim, uint32(hiddenSize/nHeads)))
+			partialRotaryFactor := cmp.Or(m.TextConfig.PartialRotaryFactor, m.TextConfig.RopeParameters.PartialRotaryFactor, float32(1.0))
+
+			// Use appropriate head count: nHeads for Q, nKVHeads for K
+			effectiveHeads := nHeads
+			if strings.Contains(name, "attn_k.") {
+				effectiveHeads = nKVHeads
+			}
+
+			permutedT := t.Clone()
+			permutedT.SetRepacker(normalToNeoXRepacker(effectiveHeads, headDim, partialRotaryFactor))
+			out = append(out, &ggml.Tensor{
+				Name:     name,
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: permutedT,
+			})
+			continue
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     name,
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (m *glmOcrModel) Replacements() []string {
+	return []string{
+		// Vision encoder
+		"model.visual.patch_embed.proj_1", "v.patch_embd_1", // Second temporal split
+		"model.visual.patch_embed.proj", "v.patch_embd",
+		"model.visual.blocks", "v.blk",
+		"model.visual.post_layernorm", "v.post_ln",
+		"model.visual.downsample", "mm.patch_merger",
+
+		// Vision attention
+		"attn.qkv", "attn_qkv",
+		"attn.proj", "attn_out",
+		"attn.q_norm", "attn_q_norm",
+		"attn.k_norm", "attn_k_norm",
+
+		// Vision norms
+		"norm1", "ln1",
+		"norm2", "ln2",
+
+		// Vision MLP
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.up_proj", "ffn_up",
+		"mlp.down_proj", "ffn_down",
+
+		// Merger (multimodal projector)
+		"model.visual.merger.proj", "mm.model.fc",
+		"model.visual.merger.post_projection_norm", "mm.post_norm",
+		"model.visual.merger.gate_proj", "mm.gate",
+		"model.visual.merger.up_proj", "mm.up",
+		"model.visual.merger.down_proj", "mm.down",
+
+		// Language model
+		"model.language_model.embed_tokens", "token_embd",
+		"model.language_model.layers", "blk",
+		"model.language_model.norm", "output_norm",
+		"lm_head", "output",
+
+		// Language model attention
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_out",
+
+		// Language model norms
+		"input_layernorm", "attn_norm",
+		"post_attention_layernorm", "ffn_norm",
+		"post_self_attn_layernorm", "post_attn_norm",
+		"post_mlp_layernorm", "post_ffn_norm",
+
+		// Language model MLP (remove mlp. prefix so ffn_* names work)
+		"mlp.gate_up_proj", "ffn_gate_up",
+		"mlp.down_proj", "ffn_down",
+	}
+}

convert/convert_lfm2.go

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

convert/convert_qwen3next.go

@@ -0,0 +1,512 @@
+package convert
+
+import (
+	"fmt"
+	"io/fs"
+	"math"
+	"slices"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type qwen3NextModel struct {
+	ModelParameters
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	HeadDim               uint32  `json:"head_dim"`
+	RopeTheta             float32 `json:"rope_theta"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+
+	// MoE config
+	NumExperts             uint32 `json:"num_experts"`
+	NumExpertsPerToken     uint32 `json:"num_experts_per_tok"`
+	NormTopkProb           bool   `json:"norm_topk_prob"`
+	MoEIntermediateSize    uint32 `json:"moe_intermediate_size"`
+	SharedExpertIntermSize uint32 `json:"shared_expert_intermediate_size"`
+
+	// Hybrid attention config
+	FullAttentionInterval uint32 `json:"full_attention_interval"`
+
+	// Linear attention (Gated Delta Net) config
+	LinearConvKernelDim uint32 `json:"linear_conv_kernel_dim"`
+	LinearKeyHeadDim    uint32 `json:"linear_key_head_dim"`
+	LinearNumKeyHeads   uint32 `json:"linear_num_key_heads"`
+	LinearNumValueHeads uint32 `json:"linear_num_value_heads"`
+	LinearValueHeadDim  uint32 `json:"linear_value_head_dim"`
+
+	// RoPE config
+	PartialRotaryFactor float32 `json:"partial_rotary_factor"`
+	RopeScaling         struct {
+		Type   string     `json:"type"`
+		Factor ropeFactor `json:"factor"`
+	} `json:"rope_scaling"`
+}
+
+var _ ModelConverter = (*qwen3NextModel)(nil)
+
+func (q *qwen3NextModel) parseMore(_ fs.FS) error {
+	if q.NumHiddenLayers == 0 {
+		return fmt.Errorf("qwen3next: num_hidden_layers must be set")
+	}
+	if q.NumAttentionHeads == 0 {
+		return fmt.Errorf("qwen3next: num_attention_heads must be set")
+	}
+	if q.NumKeyValueHeads == 0 {
+		return fmt.Errorf("qwen3next: num_key_value_heads must be set")
+	}
+	if q.HeadDim == 0 {
+		return fmt.Errorf("qwen3next: head_dim must be set")
+	}
+	if q.RopeTheta == 0 {
+		return fmt.Errorf("qwen3next: rope_theta must be set")
+	}
+	if q.PartialRotaryFactor <= 0 || q.PartialRotaryFactor > 1 {
+		return fmt.Errorf("qwen3next: partial_rotary_factor must be in (0,1], got %v", q.PartialRotaryFactor)
+	}
+	if q.LinearNumKeyHeads == 0 || q.LinearNumValueHeads == 0 || q.LinearKeyHeadDim == 0 || q.LinearValueHeadDim == 0 {
+		return fmt.Errorf("qwen3next: linear attention config must be set (linear_num_key_heads, linear_num_value_heads, linear_key_head_dim, linear_value_head_dim)")
+	}
+	if q.FullAttentionInterval == 0 {
+		return fmt.Errorf("qwen3next: full_attention_interval must be set")
+	}
+	if q.FullAttentionInterval > q.NumHiddenLayers {
+		return fmt.Errorf("qwen3next: full_attention_interval (%d) exceeds num_hidden_layers (%d)", q.FullAttentionInterval, q.NumHiddenLayers)
+	}
+
+	hasFull := false
+	for i := range q.NumHiddenLayers {
+		if (i+1)%q.FullAttentionInterval == 0 {
+			hasFull = true
+			break
+		}
+	}
+	if !hasFull {
+		return fmt.Errorf("qwen3next: head_count_kv would be all zeros (full_attention_interval=%d, num_hidden_layers=%d)", q.FullAttentionInterval, q.NumHiddenLayers)
+	}
+
+	return nil
+}
+
+func (q *qwen3NextModel) KV(t *Tokenizer) KV {
+	kv := q.ModelParameters.KV(t)
+	kv["general.architecture"] = "qwen3next"
+	kv["tokenizer.ggml.pre"] = "qwen2"
+	kv["block_count"] = q.NumHiddenLayers
+	kv["context_length"] = q.MaxPositionEmbeddings
+	kv["embedding_length"] = q.HiddenSize
+	kv["feed_forward_length"] = q.IntermediateSize
+	kv["attention.head_count"] = q.NumAttentionHeads
+	headDim := q.HeadDim
+	if headDim == 0 && q.NumAttentionHeads > 0 {
+		headDim = q.HiddenSize / q.NumAttentionHeads
+	}
+	kv["attention.key_length"] = headDim
+	kv["attention.value_length"] = headDim
+	kv["attention.layer_norm_rms_epsilon"] = q.RMSNormEPS
+	kv["rope.freq_base"] = q.RopeTheta
+
+	// RoPE dimension count (partial rotary)
+	// partial_rotary_factor = 0.25 means only 25% of head_dim uses RoPE
+	partialRotary := q.PartialRotaryFactor
+	if partialRotary > 0 && partialRotary <= 1 {
+		kv["rope.dimension_count"] = uint32(float32(headDim) * partialRotary)
+	}
+
+	// MoE config
+	if q.NumExperts > 0 {
+		kv["expert_count"] = q.NumExperts
+		kv["expert_used_count"] = q.NumExpertsPerToken
+		kv["norm_top_k_prob"] = q.NormTopkProb
+		if q.MoEIntermediateSize > 0 {
+			kv["expert_feed_forward_length"] = q.MoEIntermediateSize
+		}
+		if q.SharedExpertIntermSize > 0 {
+			kv["expert_shared_feed_forward_length"] = q.SharedExpertIntermSize
+		}
+	}
+
+	// SSM/Linear attention config
+	// d_inner = linear_value_head_dim * linear_num_value_heads
+	dInner := q.LinearValueHeadDim * q.LinearNumValueHeads
+	kv["ssm.inner_size"] = dInner
+	kv["ssm.state_size"] = q.LinearKeyHeadDim        // head_k_dim
+	kv["ssm.group_count"] = q.LinearNumKeyHeads      // num_k_heads
+	kv["ssm.time_step_rank"] = q.LinearNumValueHeads // num_v_heads
+	kv["ssm.conv_kernel"] = q.LinearConvKernelDim
+	interval := q.FullAttentionInterval
+	kv["full_attention_interval"] = interval
+
+	// Build per-layer KV head count array to identify layer types
+	// 0 = recurrent (linear attention), non-zero = full attention
+	kvHeadCounts := make([]uint32, q.NumHiddenLayers)
+	for i := range q.NumHiddenLayers {
+		// Full attention every full_attention_interval layers (starting at interval-1)
+		if interval > 0 && (i+1)%interval == 0 {
+			kvHeadCounts[i] = q.NumKeyValueHeads
+		}
+		// else stays 0 (recurrent layer)
+	}
+	kv["attention.head_count_kv"] = kvHeadCounts
+
+	// RoPE scaling
+	if q.RopeScaling.Type != "" {
+		kv["rope.scaling.type"] = q.RopeScaling.Type
+		kv["rope.scaling.factor"] = q.RopeScaling.Factor
+	}
+
+	return kv
+}
+
+func (q *qwen3NextModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+
+	// Create merges for expert tensors - stack individual experts into batched tensors
+	merges := make([]merge, q.NumHiddenLayers*3)
+	for i := range q.NumHiddenLayers {
+		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),
+		}
+	}
+
+	// Merge expert tensors
+	merged, remaining := mergeTensors(ts, merges...)
+	out = append(out, merged...)
+
+	// Process remaining tensors
+	for _, t := range remaining {
+		name := t.Name()
+		shape := t.Shape()
+
+		// Split linear_attn.in_proj_qkvz (ssm_in) into attn_qkv + attn_gate when possible
+		if strings.HasSuffix(name, ".ssm_in.weight") {
+			if qkv, gate, ok := q.splitQKVZTensor(t); ok {
+				out = append(out, qkv, gate)
+				continue
+			}
+			panic(fmt.Sprintf("qwen3next: failed to split %s into attn_qkv/attn_gate (shape=%v)", name, shape))
+		}
+
+		switch {
+		// Add 1 to norm weights (except ssm_norm which is linear_attn.norm)
+		// This matches the Python converter behavior for qwen3next
+		case strings.HasSuffix(name, "_norm.weight") && !strings.HasSuffix(name, ".ssm_norm.weight"):
+			t.SetRepacker(q.addOne)
+			out = append(out, &ggml.Tensor{
+				Name:     name,
+				Kind:     t.Kind(),
+				Shape:    slices.Clone(shape),
+				WriterTo: t,
+			})
+
+		// Handle linear attention A_log -> ssm_a (negate and exp)
+		// Note: name has already been transformed by Replacements at this point
+		case strings.HasSuffix(name, ".ssm_a"):
+			t.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
+				// Compute -exp(A_log)
+				result := make([]float32, len(data))
+				for i, v := range data {
+					// -exp(v)
+					result[i] = -float32(math.Exp(float64(v)))
+				}
+				return result, nil
+			})
+			out = append(out, &ggml.Tensor{
+				Name:     name,
+				Kind:     t.Kind(),
+				Shape:    slices.Clone(shape),
+				WriterTo: t,
+			})
+
+		// Squeeze conv1d weights: [1, D, K] or [D, 1, K] -> [D, K]
+		case strings.HasSuffix(name, ".ssm_conv1d.weight"):
+			newShape := slices.Clone(shape)
+			if len(shape) == 3 {
+				if shape[0] == 1 {
+					// [1, D, K] -> [D, K]
+					newShape = []uint64{shape[1], shape[2]}
+				} else if shape[1] == 1 {
+					// [D, 1, K] -> [D, K]
+					newShape = []uint64{shape[0], shape[2]}
+				}
+			}
+			out = append(out, &ggml.Tensor{
+				Name:     name,
+				Kind:     t.Kind(),
+				Shape:    newShape,
+				WriterTo: t,
+			})
+		// Squeeze shared expert gate: [D, 1] or [1, D] -> [D]
+		case strings.HasSuffix(name, ".ffn_gate_inp_shexp.weight"):
+			newShape := slices.Clone(shape)
+			if len(shape) == 2 {
+				if shape[0] == 1 && shape[1] > 1 {
+					newShape = []uint64{shape[1]}
+				} else if shape[1] == 1 && shape[0] > 1 {
+					newShape = []uint64{shape[0]}
+				}
+			}
+			out = append(out, &ggml.Tensor{
+				Name:     name,
+				Kind:     t.Kind(),
+				Shape:    newShape,
+				WriterTo: t,
+			})
+
+		default:
+			out = append(out, &ggml.Tensor{
+				Name:     name,
+				Kind:     t.Kind(),
+				Shape:    slices.Clone(shape),
+				WriterTo: t,
+			})
+		}
+	}
+
+	return out
+}
+
+type qkvzSplitSpec struct {
+	hidden    int
+	headKDim  int
+	headVDim  int
+	numKHeads int
+	numVHeads int
+	qkvzDim   int
+	qkvOut    int
+	gateOut   int
+}
+
+func (q *qwen3NextModel) qkvzSpec(shape []uint64) (qkvzSplitSpec, bool) {
+	if len(shape) != 2 {
+		return qkvzSplitSpec{}, false
+	}
+
+	numKHeads := int(q.LinearNumKeyHeads)
+	numVHeads := int(q.LinearNumValueHeads)
+	headKDim := int(q.LinearKeyHeadDim)
+	headVDim := int(q.LinearValueHeadDim)
+	if numKHeads == 0 || numVHeads == 0 || headKDim == 0 || headVDim == 0 {
+		return qkvzSplitSpec{}, false
+	}
+	if numVHeads%numKHeads != 0 {
+		return qkvzSplitSpec{}, false
+	}
+
+	hidden := int(shape[1])
+	vPerHead := headVDim * (numVHeads / numKHeads)
+	qkvzDim := 2*headKDim + 2*vPerHead
+	expectedOut := qkvzDim * numKHeads
+	if int(shape[0]) != expectedOut {
+		return qkvzSplitSpec{}, false
+	}
+
+	return qkvzSplitSpec{
+		hidden:    hidden,
+		headKDim:  headKDim,
+		headVDim:  headVDim,
+		numKHeads: numKHeads,
+		numVHeads: numVHeads,
+		qkvzDim:   qkvzDim,
+		qkvOut:    2*headKDim*numKHeads + headVDim*numVHeads,
+		gateOut:   headVDim * numVHeads,
+	}, true
+}
+
+func (q *qwen3NextModel) splitQKVZTensor(t Tensor) (*ggml.Tensor, *ggml.Tensor, bool) {
+	spec, ok := q.qkvzSpec(t.Shape())
+	if !ok {
+		return nil, nil, false
+	}
+
+	qkvTensor := t.Clone()
+	qkvTensor.SetRepacker(q.repackQKVZ(spec, false))
+
+	gateTensor := t.Clone()
+	gateTensor.SetRepacker(q.repackQKVZ(spec, true))
+
+	qkvName := strings.Replace(t.Name(), "ssm_in", "attn_qkv", 1)
+	gateName := strings.Replace(t.Name(), "ssm_in", "attn_gate", 1)
+
+	return &ggml.Tensor{
+			Name:     qkvName,
+			Kind:     t.Kind(),
+			Shape:    []uint64{uint64(spec.qkvOut), uint64(spec.hidden)},
+			WriterTo: qkvTensor,
+		}, &ggml.Tensor{
+			Name:     gateName,
+			Kind:     t.Kind(),
+			Shape:    []uint64{uint64(spec.gateOut), uint64(spec.hidden)},
+			WriterTo: gateTensor,
+		}, true
+}
+
+func (q *qwen3NextModel) repackQKVZ(spec qkvzSplitSpec, extractGate bool) Repacker {
+	vPerHead := spec.headVDim * (spec.numVHeads / spec.numKHeads)
+
+	return func(_ string, data []float32, shape []uint64) ([]float32, error) {
+		dims := make([]int, len(shape))
+		for i := range shape {
+			dims[i] = int(shape[i])
+		}
+
+		var tt tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+		var err error
+
+		// Convert to [hidden, out_features] layout for slicing
+		tt, err = tensor.Transpose(tt, 1, 0)
+		if err != nil {
+			return nil, err
+		}
+		tt = tensor.Materialize(tt)
+
+		if err := tt.Reshape(spec.hidden, spec.numKHeads, spec.qkvzDim); err != nil {
+			return nil, err
+		}
+
+		offset := 0
+		qSlice, err := tt.Slice(nil, nil, tensor.S(offset, offset+spec.headKDim))
+		if err != nil {
+			return nil, err
+		}
+		offset += spec.headKDim
+		kSlice, err := tt.Slice(nil, nil, tensor.S(offset, offset+spec.headKDim))
+		if err != nil {
+			return nil, err
+		}
+		offset += spec.headKDim
+		vSlice, err := tt.Slice(nil, nil, tensor.S(offset, offset+vPerHead))
+		if err != nil {
+			return nil, err
+		}
+		offset += vPerHead
+		zSlice, err := tt.Slice(nil, nil, tensor.S(offset, offset+vPerHead))
+		if err != nil {
+			return nil, err
+		}
+
+		qMat := tensor.Materialize(qSlice).(*tensor.Dense)
+		kMat := tensor.Materialize(kSlice).(*tensor.Dense)
+		vMat := tensor.Materialize(vSlice).(*tensor.Dense)
+		zMat := tensor.Materialize(zSlice).(*tensor.Dense)
+
+		if err := qMat.Reshape(spec.hidden, spec.numKHeads*spec.headKDim); err != nil {
+			return nil, err
+		}
+		if err := kMat.Reshape(spec.hidden, spec.numKHeads*spec.headKDim); err != nil {
+			return nil, err
+		}
+		if err := vMat.Reshape(spec.hidden, spec.numKHeads*vPerHead); err != nil {
+			return nil, err
+		}
+		if err := zMat.Reshape(spec.hidden, spec.numKHeads*vPerHead); err != nil {
+			return nil, err
+		}
+
+		var out tensor.Tensor
+		if extractGate {
+			out = zMat
+		} else {
+			out, err = tensor.Concat(1, qMat, kMat, vMat)
+			if err != nil {
+				return nil, err
+			}
+		}
+
+		out = tensor.Materialize(out)
+		out, err = tensor.Transpose(out, 1, 0)
+		if err != nil {
+			return nil, err
+		}
+		out = tensor.Materialize(out)
+
+		if err := out.Reshape(out.Shape().TotalSize()); err != nil {
+			return nil, err
+		}
+
+		return native.VectorF32(out.(*tensor.Dense))
+	}
+}
+
+// addOne adds 1.0 to all elements in the tensor (for norm weights)
+func (*qwen3NextModel) addOne(_ string, data []float32, shape []uint64) ([]float32, error) {
+	n := tensor.New(tensor.WithShape(int(shape[0])), tensor.WithBacking(data))
+	ones := tensor.Ones(tensor.Float32, int(shape[0]))
+
+	n, err := n.Add(ones)
+	if err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 0)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}
+
+func (q *qwen3NextModel) Replacements() []string {
+	return []string{
+		// Embeddings and output
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+
+		// Layer norms
+		"input_layernorm", "attn_norm",
+		"post_attention_layernorm", "post_attention_norm",
+
+		// Full attention (self_attn)
+		"self_attn.q_proj", "attn_q",
+		"self_attn.q_norm", "attn_q_norm",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.k_norm", "attn_k_norm",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+
+		// Linear attention (Gated Delta Net)
+		"linear_attn.in_proj_qkvz", "ssm_in",
+		"linear_attn.in_proj_ba", "ssm_ba",
+		"linear_attn.conv1d", "ssm_conv1d",
+		"linear_attn.dt_bias", "ssm_dt",
+		"linear_attn.dt_proj", "ssm_dt",
+		"linear_attn.A_log", "ssm_a",
+		"linear_attn.norm", "ssm_norm",
+		"linear_attn.out_proj", "ssm_out",
+
+		// MoE (experts are stacked via mergeTensors, not replaced here)
+		"mlp.gate.weight", "ffn_gate_inp.weight",
+		"mlp.shared_expert.down_proj", "ffn_down_shexp",
+		"mlp.shared_expert.gate_proj", "ffn_gate_shexp",
+		"mlp.shared_expert.up_proj", "ffn_up_shexp",
+		"mlp.shared_expert_gate", "ffn_gate_inp_shexp",
+
+		// Dense FFN (if any layers use it)
+		"mlp.down_proj", "ffn_down",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.up_proj", "ffn_up",
+	}
+}

convert/reader.go

@@ -40,6 +40,8 @@ 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") ||
+		strings.HasSuffix(t.name, ".ssm_conv1d.weight") || // SSM conv kernel must be F32 for Metal
 		t.name == "token_types.weight" ||
 		t.name == "v.positional_embedding_vlm" ||
 		t.name == "v.tile_position_embd.weight" ||

convert/reader_safetensors.go

@@ -99,6 +99,8 @@ func (st safetensor) Kind() uint32 {
 	if st.dtype == "BF16" &&
 		!strings.HasPrefix(st.name, "v.") &&
 		!strings.HasPrefix(st.name, "s.") &&
+		!strings.HasPrefix(st.name, "mm.") &&
+		!strings.Contains(st.name, "ffn_gate_inp_shexp.weight") &&
 		kind != tensorKindFP32 {
 		kind = tensorKindBF16
 	}

docs/api/anthropic-compatibility.mdx

@@ -4,16 +4,6 @@ title: Anthropic compatibility
 
 Ollama provides compatibility with the [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) to help connect existing applications to Ollama, including tools like Claude Code.
 
-## Recommended models
-
-For coding use cases, models like `glm-4.7:cloud`, `minimax-m2.1:cloud`, and `qwen3-coder` are recommended.
-
-Pull a model before use:
-```shell
-ollama pull qwen3-coder
-ollama pull glm-4.7:cloud
-```
-
 ## Usage
 
 ### Environment variables
@@ -22,8 +12,8 @@ To use Ollama with tools that expect the Anthropic API (like Claude Code), set t
 
 ```shell
 export ANTHROPIC_AUTH_TOKEN=ollama  # required but ignored
+export ANTHROPIC_API_KEY="" # required but ignored
 export ANTHROPIC_BASE_URL=http://localhost:11434
-export ANTHROPIC_API_KEY=ollama  # required but ignored
 ```
 
 ### Simple `/v1/messages` example
@@ -245,10 +235,41 @@ curl -X POST http://localhost:11434/v1/messages \
 
 ## Using with Claude Code
 
-[Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
+[Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend. 
+
+### Recommended models
+
+For coding use cases, models like `glm-4.7`, `minimax-m2.1`, and `qwen3-coder` are recommended.
+
+Download a model before use:
 
 ```shell
-ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+ollama pull qwen3-coder
+```
+> Note: Qwen 3 coder is a 30B parameter model requiring at least 24GB of VRAM to run smoothly. More is required for longer context lengths. 
+
+```shell
+ollama pull glm-4.7:cloud
+```
+
+### Quick setup
+
+```shell
+ollama launch claude
+```
+
+This will prompt you to select a model, configure Claude Code automatically, and launch it. To configure without launching:
+
+```shell
+ollama launch claude --config
+```
+
+### Manual setup
+
+Set the environment variables and run Claude Code:
+
+```shell
+ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY="" claude --model qwen3-coder
 ```
 
 Or set the environment variables in your shell profile:
@@ -256,19 +277,13 @@ 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
+export ANTHROPIC_API_KEY=""
 ```
 
 Then run Claude Code with any Ollama model:
 
 ```shell
-# Local models
 claude --model qwen3-coder
-claude --model gpt-oss:20b
-
-# Cloud models
-claude --model glm-4.7:cloud
-claude --model minimax-m2.1:cloud
 ```
 
 ## Endpoints

docs/cli.mdx

@@ -8,6 +8,47 @@ title: CLI Reference
 ollama run gemma3
 ```
 
+### Launch integrations
+
+```
+ollama launch
+```
+
+Configure and launch external applications to use Ollama models. This provides an interactive way to set up and start integrations with supported apps.
+
+#### Supported integrations
+
+- **OpenCode** - Open-source coding assistant
+- **Claude Code** - Anthropic's agentic coding tool
+- **Codex** - OpenAI's coding assistant
+- **Droid** - Factory's AI coding agent
+
+#### Examples
+
+Launch an integration interactively:
+
+```
+ollama launch
+```
+
+Launch a specific integration:
+
+```
+ollama launch claude
+```
+
+Launch with a specific model:
+
+```
+ollama launch claude --model qwen3-coder
+```
+
+Configure without launching:
+
+```
+ollama launch droid --config
+```
+
 #### Multiline input
 
 For multiline input, you can wrap text with `"""`:

docs/cloud.mdx

@@ -3,8 +3,6 @@ title: Cloud
 sidebarTitle: Cloud
 ---
 
-<Info>Ollama's cloud is currently in preview.</Info>
-
 ## Cloud Models
 
 Ollama's cloud models are a new kind of model in Ollama that can run without a powerful GPU. Instead, cloud models are automatically offloaded to Ollama's cloud service while offering the same capabilities as local models, making it possible to keep using your local tools while running larger models that wouldn't fit on a personal computer.

docs/context-length.mdx

@@ -5,10 +5,13 @@ title: Context length
 Context length is the maximum number of tokens that the model has access to in memory.  
 
 <Note>
-  The default context length in Ollama is 4096 tokens.
+  Ollama defaults to the following context lengths based on VRAM:
+    - < 24 GiB VRAM: 4k context
+    - 24-48 GiB VRAM: 32k context
+    - &gt;= 48 GiB VRAM: 256k context
 </Note>
 
-Tasks which require large context like web search, agents, and coding tools should be set to at least 32000 tokens.
+Tasks which require large context like web search, agents, and coding tools should be set to at least 64000 tokens.
 
 ## Setting context length
 
@@ -24,7 +27,7 @@ Change the slider in the Ollama app under settings to your desired context lengt
 ### CLI
 If editing the context length for Ollama is not possible, the context length can also be updated when serving Ollama.  
 ```
-OLLAMA_CONTEXT_LENGTH=32000 ollama serve
+OLLAMA_CONTEXT_LENGTH=64000 ollama serve
 ```
 
 ### Check allocated context length and model offloading

docs/docs.json

@@ -71,6 +71,10 @@
     {
       "source": "/api",
       "destination": "/api/introduction"
+    },
+    {
+      "source": "/integrations/clawdbot",
+      "destination": "/integrations/openclaw"
     }
   ],
   "navigation": {
@@ -101,19 +105,52 @@
           {
             "group": "Integrations",
             "pages": [
-              "/integrations/claude-code",
-              "/integrations/vscode",
-              "/integrations/jetbrains",
-              "/integrations/codex",
-              "/integrations/cline",
-              "/integrations/droid",
-              "/integrations/goose",
-              "/integrations/zed",
-              "/integrations/roo-code",
-              "/integrations/n8n",
-              "/integrations/xcode",
-              "/integrations/onyx",
-              "/integrations/marimo"
+              "/integrations/index",
+              {
+                "group": "Coding",
+                "pages": [
+                  "/integrations/claude-code",
+                  "/integrations/codex",
+                  "/integrations/opencode",
+                  "/integrations/droid",
+                  "/integrations/goose"
+                ]
+              },
+              {
+                "group": "Assistants",
+                "pages": [
+                  "/integrations/openclaw"
+                ]
+              },
+              {
+                "group": "IDEs & Editors",
+                "pages": [
+                  "/integrations/cline",
+                  "/integrations/jetbrains",
+                  "/integrations/roo-code",
+                  "/integrations/vscode",
+                  "/integrations/xcode",
+                  "/integrations/zed"
+                ]
+              },
+              {
+                "group": "Chat & RAG",
+                "pages": [
+                  "/integrations/onyx"
+                ]
+              },
+              {
+                "group": "Automation",
+                "pages": [
+                  "/integrations/n8n"
+                ]
+              },
+              {
+                "group": "Notebooks",
+                "pages": [
+                  "/integrations/marimo"
+                ]
+              }
             ]
           },
           {

docs/faq.mdx

@@ -14,11 +14,11 @@ curl -fsSL https://ollama.com/install.sh | sh
 
 ## How can I view the logs?
 
-Review the [Troubleshooting](./troubleshooting) docs for more about using logs.
+Review the [Troubleshooting](./troubleshooting.mdx) docs for more about using logs.
 
 ## Is my GPU compatible with Ollama?
 
-Please refer to the [GPU docs](./gpu).
+Please refer to the [GPU docs](./gpu.mdx).
 
 ## How can I specify the context window size?
 
@@ -66,7 +66,7 @@ llama3:70b  bcfb190ca3a7  42 GB   100% GPU    4 minutes from now
 ```
 </Info>
 
-The `Processor` column will show which memory the model was loaded in to:
+The `Processor` column will show which memory the model was loaded into:
 
 - `100% GPU` means the model was loaded entirely into the GPU
 - `100% CPU` means the model was loaded entirely in system memory
@@ -158,7 +158,7 @@ docker run -d -e HTTPS_PROXY=https://my.proxy.example.com -p 11434:11434 ollama-
 
 ## Does Ollama send my prompts and answers back to ollama.com?
 
-No. Ollama runs locally, and conversation data does not leave your machine.
+Ollama runs locally. We don't see your prompts or data when you run locally. When using cloud-hosted models, we process your prompts and responses to provide the service but do not store or log that content and never train on it. We collect basic account info and limited usage metadata to provide the service that does not include prompt or response content. We don't sell your data. You can delete your account anytime.
 
 ## How can I expose Ollama on my network?
 
@@ -183,7 +183,7 @@ server {
 
 ## How can I use Ollama with ngrok?
 
-Ollama can be accessed using a range of tools for tunneling tools. For example with Ngrok:
+Ollama can be accessed using a range of tunneling apps. For example with Ngrok:
 
 ```shell
 ngrok http 11434 --host-header="localhost:11434"
@@ -240,7 +240,7 @@ GPU acceleration is not available for Docker Desktop in macOS due to the lack of
 
 This can impact both installing Ollama, as well as downloading models.
 
-Open `Control Panel > Networking and Internet > View network status and tasks` and click on `Change adapter settings` on the left panel. Find the `vEthernel (WSL)` adapter, right click and select `Properties`.
+Open `Control Panel > Networking and Internet > View network status and tasks` and click on `Change adapter settings` on the left panel. Find the `vEthernet (WSL)` adapter, right click and select `Properties`.
 Click on `Configure` and open the `Advanced` tab. Search through each of the properties until you find `Large Send Offload Version 2 (IPv4)` and `Large Send Offload Version 2 (IPv6)`. _Disable_ both of these
 properties.
 
@@ -299,7 +299,7 @@ The `keep_alive` API parameter with the `/api/generate` and `/api/chat` API endp
 
 ## How do I manage the maximum number of requests the Ollama server can queue?
 
-If too many requests are sent to the server, it will respond with a 503 error indicating the server is overloaded. You can adjust how many requests may be queue by setting `OLLAMA_MAX_QUEUE`.
+If too many requests are sent to the server, it will respond with a 503 error indicating the server is overloaded. You can adjust how many requests may be queued by setting `OLLAMA_MAX_QUEUE`.
 
 ## How does Ollama handle concurrent requests?
 
@@ -312,10 +312,10 @@ Parallel request processing for a given model results in increasing the context
 The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms:
 
 - `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory. The default is 3 \* the number of GPUs or 3 for CPU inference.
-- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time. The default will auto-select either 4 or 1 based on available memory.
+- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time, default 1.  Required RAM will scale by `OLLAMA_NUM_PARALLEL` * `OLLAMA_CONTEXT_LENGTH`.
 - `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512
 
-Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting. Once ROCm v6.2 is available, Windows Radeon will follow the defaults above. You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.
+Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting. Once ROCm v6.2 is available, Windows Radeon will follow the defaults above. You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPU's VRAM.
 
 ## How does Ollama load models on multiple GPUs?
 
@@ -382,7 +382,7 @@ ollama signin
   Replace &lt;username&gt; with your actual Windows user name.
 </Note>
 
-## How can I stop Ollama from starting when I login to my computer
+## How can I stop Ollama from starting when I login to my computer?
 
 Ollama for Windows and macOS register as a login item during installation.  You can disable this if you prefer not to have Ollama automatically start.  Ollama will respect this setting across upgrades, unless you uninstall the application.
 
@@ -390,4 +390,4 @@ Ollama for Windows and macOS register as a login item during installation.  You
 - In `Task Manager` go to the `Startup apps` tab, search for `ollama` then click `Disable`
 
 **MacOS**
-- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.
+- Open `Settings` and search for "Login Items", find the `Ollama` entry under `Allow in the Background`, then click the slider to disable.

docs/gpu.mdx

@@ -10,6 +10,7 @@ Check your compute compatibility to see if your card is supported:
 
 | Compute Capability | Family              | Cards                                                                                                                          |
 | ------------------ | ------------------- | ------------------------------------------------------------------------------------------------------------------------------ |
+| 12.1               | NVIDIA              | `GB10 (DGX Spark)`                                                                                                             |
 | 12.0               | GeForce RTX 50xx    | `RTX 5060` `RTX 5060 Ti` `RTX 5070` `RTX 5070 Ti` `RTX 5080` `RTX 5090`                                                        |
 |                    | NVIDIA Professional | `RTX PRO 4000 Blackwell` `RTX PRO 4500 Blackwell` `RTX PRO 5000 Blackwell` `RTX PRO 6000 Blackwell`                            |
 | 9.0                | NVIDIA              | `H200` `H100`                                                                                                                  |
@@ -163,4 +164,4 @@ To select specific Vulkan GPU(s), you can set the environment variable
 `GGML_VK_VISIBLE_DEVICES` to one or more numeric IDs on the Ollama server as
 described in the [FAQ](faq#how-do-i-configure-ollama-server). If you
 encounter any problems with Vulkan based GPUs, you can disable all Vulkan GPUs
-by setting `GGML_VK_VISIBLE_DEVICES=-1` 
+by setting `GGML_VK_VISIBLE_DEVICES=-1` 

docs/import.mdx

@@ -134,22 +134,12 @@ success
 
 ### Supported Quantizations
 
-- `q4_0`
-- `q4_1`
-- `q5_0`
-- `q5_1`
 - `q8_0`
 
 #### K-means Quantizations
 
-- `q3_K_S`
-- `q3_K_M`
-- `q3_K_L`
 - `q4_K_S`
 - `q4_K_M`
-- `q5_K_S`
-- `q5_K_M`
-- `q6_K`
 
 ## Sharing your model on ollama.com
 

docs/index.mdx

@@ -9,7 +9,7 @@ sidebarTitle: Welcome
 
 <CardGroup cols={2}>
   <Card title="Quickstart" icon="rocket" href="/quickstart">
-    Get up and running with your first model
+    Get up and running with your first model or integrate Ollama with your favorite tools
   </Card>
   <Card
     title="Download Ollama"

docs/integrations/claude-code.mdx

@@ -4,7 +4,7 @@ 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.
+Open models can be used with Claude Code through Ollama's Anthropic-compatible API, enabling you to use models such as `glm-4.7`, `qwen3-coder`, `gpt-oss`. 
 
 ![Claude Code with Ollama](https://files.ollama.com/claude-code.png)
 
@@ -26,12 +26,27 @@ irm https://claude.ai/install.ps1 | iex
 
 ## Usage with Ollama
 
+### Quick setup
+
+```shell
+ollama launch claude
+```
+
+To configure without launching:
+
+```shell
+ollama launch claude --config
+```
+
+### Manual setup
+
 Claude Code connects to Ollama using the Anthropic-compatible API.
 
 1. Set the environment variables:
 
 ```shell
 export ANTHROPIC_AUTH_TOKEN=ollama
+export ANTHROPIC_API_KEY=""
 export ANTHROPIC_BASE_URL=http://localhost:11434
 ```
 
@@ -44,35 +59,17 @@ claude --model gpt-oss:20b
 Or run with environment variables inline:
 
 ```shell
-ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 claude --model gpt-oss:20b
+ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY="" 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
-2. Set the environment variables:
-
-```shell
-export ANTHROPIC_BASE_URL=https://ollama.com
-export ANTHROPIC_API_KEY=<your-api-key>
-```
-
-3. Run Claude Code with a cloud model:
-
-```shell
-claude --model glm-4.7:cloud
-```
+**Note:** Claude Code requires a large context window. We recommend at least 64k tokens. See the [context length documentation](/context-length) for how to adjust context length in Ollama.
 
 ## Recommended Models
 
-### Cloud models
-- `glm-4.7:cloud` - High-performance cloud model
-- `minimax-m2.1:cloud` - Fast cloud model
-- `qwen3-coder:480b` - Large coding model
+- `qwen3-coder` 
+- `glm-4.7`
+- `gpt-oss:20b`
+- `gpt-oss:120b`
+
+Cloud models are also available at [ollama.com/search?c=cloud](https://ollama.com/search?c=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/codex.mdx

@@ -13,7 +13,21 @@ npm install -g @openai/codex
 
 ## Usage with Ollama
 
-<Note>Codex requires a larger context window. It is recommended to use a context window of at least 32K tokens.</Note>
+<Note>Codex requires a larger context window. It is recommended to use a context window of at least 64k tokens.</Note>
+
+### Quick setup
+
+```
+ollama launch codex
+```
+
+To configure without launching:
+
+```shell
+ollama launch codex --config
+```
+
+### Manual setup
 
 To use `codex` with Ollama, use the `--oss` flag:
 

docs/integrations/droid.mdx

@@ -11,10 +11,24 @@ Install the [Droid CLI](https://factory.ai/):
 curl -fsSL https://app.factory.ai/cli | sh
 ```
 
-<Note>Droid requires a larger context window. It is recommended to use a context window of at least 32K tokens. See [Context length](/context-length) for more information.</Note>
+<Note>Droid requires a larger context window. It is recommended to use a context window of at least 64k tokens. See [Context length](/context-length) for more information.</Note>
 
 ## Usage with Ollama
 
+### Quick setup
+
+```bash
+ollama launch droid
+```
+
+To configure without launching:
+
+```shell
+ollama launch droid --config
+```
+
+### Manual setup
+
 Add a local configuration block to `~/.factory/config.json`:
 
 ```json
@@ -73,4 +87,4 @@ Add the cloud configuration block to `~/.factory/config.json`:
    }
    ```
 
-Run `droid` in a new terminal to load the new settings.
+Run `droid` in a new terminal to load the new settings.

docs/integrations/index.mdx

@@ -0,0 +1,50 @@
+---
+title: Overview
+---
+
+Ollama integrates with a wide range of tools.
+
+## Coding Agents
+
+Coding assistants that can read, modify, and execute code in your projects.
+
+- [Claude Code](/integrations/claude-code)
+- [Codex](/integrations/codex)
+- [OpenCode](/integrations/opencode)
+- [Droid](/integrations/droid)
+- [Goose](/integrations/goose)
+
+## Assistants
+
+AI assistants that help with everyday tasks.
+
+- [OpenClaw](/integrations/openclaw)
+
+## IDEs & Editors
+
+Native integrations for popular development environments.
+
+- [VS Code](/integrations/vscode)
+- [Cline](/integrations/cline)
+- [Roo Code](/integrations/roo-code)
+- [JetBrains](/integrations/jetbrains)
+- [Xcode](/integrations/xcode)
+- [Zed](/integrations/zed)
+
+## Chat & RAG
+
+Chat interfaces and retrieval-augmented generation platforms.
+
+- [Onyx](/integrations/onyx)
+
+## Automation
+
+Workflow automation platforms with AI integration.
+
+- [n8n](/integrations/n8n)
+
+## Notebooks
+
+Interactive computing environments with AI capabilities.
+
+- [marimo](/integrations/marimo)

docs/integrations/openclaw.mdx

@@ -0,0 +1,50 @@
+---
+title: OpenClaw
+---
+
+OpenClaw is a personal AI assistant that runs on your own devices. It bridges messaging services (WhatsApp, Telegram, Slack, Discord, iMessage, and more) to AI coding agents through a centralized gateway.
+
+## Install
+
+Install [OpenClaw](https://openclaw.ai/) 
+
+```bash
+npm install -g openclaw@latest
+```
+
+Then run the onboarding wizard:
+
+```bash
+openclaw onboard --install-daemon
+```
+
+<Note>OpenClaw requires a larger context window. It is recommended to use a context window of at least 64k tokens. See [Context length](/context-length) for more information.</Note>
+
+## Usage with Ollama
+
+### Quick setup
+
+```bash
+ollama launch openclaw
+```
+
+<Note>Previously known as Clawdbot. `ollama launch clawdbot` still works as an alias.</Note>
+
+This configures OpenClaw to use Ollama and starts the gateway.
+If the gateway is already running, no changes need to be made as the gateway will auto-reload the changes. 
+
+
+To configure without launching:
+
+```shell
+ollama launch openclaw --config
+```
+
+## Recommended Models
+
+- `qwen3-coder`
+- `glm-4.7`
+- `gpt-oss:20b`
+- `gpt-oss:120b`
+
+Cloud models are also available at [ollama.com/search?c=cloud](https://ollama.com/search?c=cloud).

docs/integrations/opencode.mdx

@@ -0,0 +1,106 @@
+---
+title: OpenCode
+---
+
+OpenCode is an open-source AI coding assistant that runs in your terminal.
+
+## Install
+
+Install the [OpenCode CLI](https://opencode.ai):
+
+```bash
+curl -fsSL https://opencode.ai/install | bash
+```
+
+<Note>OpenCode requires a larger context window. It is recommended to use a context window of at least 64k tokens. See [Context length](/context-length) for more information.</Note>
+
+## Usage with Ollama
+
+### Quick setup
+
+```bash
+ollama launch opencode
+```
+
+To configure without launching:
+
+```shell
+ollama launch opencode --config
+```
+
+### Manual setup
+
+Add a configuration block to `~/.config/opencode/opencode.json`:
+
+```json
+{
+  "$schema": "https://opencode.ai/config.json",
+  "provider": {
+    "ollama": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Ollama",
+      "options": {
+        "baseURL": "http://localhost:11434/v1"
+      },
+      "models": {
+        "qwen3-coder": {
+          "name": "qwen3-coder"
+        }
+      }
+    }
+  }
+}
+```
+
+## Cloud Models
+
+`glm-4.7:cloud` is the recommended model for use with OpenCode.
+
+Add the cloud configuration to `~/.config/opencode/opencode.json`:
+
+```json
+{
+  "$schema": "https://opencode.ai/config.json",
+  "provider": {
+    "ollama": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Ollama",
+      "options": {
+        "baseURL": "http://localhost:11434/v1"
+      },
+      "models": {
+        "glm-4.7:cloud": {
+          "name": "glm-4.7:cloud"
+        }
+      }
+    }
+  }
+}
+```
+
+## Connecting to ollama.com
+
+1. Create an [API key](https://ollama.com/settings/keys) from ollama.com and export it as `OLLAMA_API_KEY`.
+2. Update `~/.config/opencode/opencode.json` to point to ollama.com:
+
+```json
+{
+  "$schema": "https://opencode.ai/config.json",
+  "provider": {
+    "ollama": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Ollama Cloud",
+      "options": {
+        "baseURL": "https://ollama.com/v1"
+      },
+      "models": {
+        "glm-4.7:cloud": {
+          "name": "glm-4.7:cloud"
+        }
+      }
+    }
+  }
+}
+```
+
+Run `opencode` in a new terminal to load the new settings.

docs/quickstart.mdx

@@ -2,7 +2,7 @@
 title: Quickstart
 ---
 
-This quickstart will walk your through running your first model with Ollama. To get started, download Ollama on macOS, Windows or Linux.
+Ollama is available on macOS, Windows, and Linux.
 
 <a
   href="https://ollama.com/download"
@@ -12,92 +12,48 @@ This quickstart will walk your through running your first model with Ollama. To
   Download Ollama
 </a>
 
-## Run a model
+## Get Started
 
-<Tabs>
-  <Tab title="CLI">
-    Open a terminal and run the command:
+Run `ollama` in your terminal to open the interactive menu:
 
-    ```
-    ollama run gemma3
-    ```
+```sh
+ollama
+```
 
-  </Tab>
-  <Tab title="cURL">
-    ```
-    ollama pull gemma3
-    ```
+Navigate with `↑/↓`, press `enter` to launch, `→` to change model, and `esc` to quit.
 
-    Lastly, chat with the model:
+The menu provides quick access to:
+- **Run a model** - Start an interactive chat
+- **Launch tools** - Claude Code, Codex, OpenClaw, and more
+- **Additional integrations** - Available under "More..."
 
-    ```shell
-    curl http://localhost:11434/api/chat -d '{
-      "model": "gemma3",
-      "messages": [{
-        "role": "user",
-        "content": "Hello there!"
-      }],
-      "stream": false
-    }'
-    ```
+## Coding
 
-  </Tab>
-  <Tab title="Python">
-    Start by downloading a model:
+Launch coding tools with Ollama models:
 
-    ```
-    ollama pull gemma3
-    ```
+```sh
+ollama launch claude
+```
 
-    Then install Ollama's Python library:
+```sh
+ollama launch codex
+```
 
-    ```
-    pip install ollama
-    ```
+```sh
+ollama launch opencode
+```
 
-    Lastly, chat with the model:
+See [integrations](/integrations) for all supported tools.
 
-    ```python
-    from ollama import chat
-    from ollama import ChatResponse
+## API
 
-    response: ChatResponse = chat(model='gemma3', messages=[
-      {
-        'role': 'user',
-        'content': 'Why is the sky blue?',
-      },
-    ])
-    print(response['message']['content'])
-    # or access fields directly from the response object
-    print(response.message.content)
-    ```
+Use the [API](/api) to integrate Ollama into your applications:
 
-  </Tab>
-  <Tab title="JavaScript">
-    Start by downloading a model:
+```sh
+curl http://localhost:11434/api/chat -d '{
+  "model": "gemma3",
+  "messages": [{ "role": "user", "content": "Hello!" }]
+}'
+```
 
-    ```
-    ollama pull gemma3
-    ```
-
-    Then install the Ollama JavaScript library:
-    ```
-    npm i ollama
-    ```
-
-    Lastly, chat with the model:
-
-    ```shell
-    import ollama from 'ollama'
-
-    const response = await ollama.chat({
-      model: 'gemma3',
-      messages: [{ role: 'user', content: 'Why is the sky blue?' }],
-    })
-    console.log(response.message.content)
-    ```
-
-  </Tab>
-</Tabs>
-
-See a full list of available models [here](https://ollama.com/models).
+See the [API documentation](/api) for Python, JavaScript, and other integrations.

envconfig/config.go

@@ -201,7 +201,7 @@ var (
 	// Enable the new Ollama engine
 	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
-	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
+	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 0)
 	// Auth enables authentication between the Ollama client and server
 	UseAuth = Bool("OLLAMA_AUTH")
 	// Enable Vulkan backend
@@ -290,7 +290,7 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowedOrigins(), "A comma separated list of allowed origins"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
-		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4096)"},
+		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4k/32k/256k based on VRAM)"},
 		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
 		"OLLAMA_REMOTES":           {"OLLAMA_REMOTES", Remotes(), "Allowed hosts for remote models (default \"ollama.com\")"},
 

envconfig/config_test.go

@@ -282,7 +282,7 @@ func TestVar(t *testing.T) {
 
 func TestContextLength(t *testing.T) {
 	cases := map[string]uint{
-		"":     4096,
+		"":     0,
 		"2048": 2048,
 	}
 

fs/ggml/ggml.go

@@ -268,7 +268,11 @@ func (kv KV) OllamaEngineRequired() bool {
 		"olmo3",
 		"qwen25vl",
 		"qwen3", "qwen3moe",
+		"qwen3next",
 		"qwen3vl", "qwen3vlmoe",
+		"glm4moelite",
+		"glmocr",
+		"lfm2",
 	}, kv.Architecture())
 }
 
@@ -856,10 +860,14 @@ func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
 		"bert",
 		"gemma3",
+		"glm4moelite",
+		"glmocr",
 		"gptoss", "gpt-oss",
+		"lfm2",
 		"mistral3",
 		"olmo3",
 		"qwen3", "qwen3moe",
+		"qwen3next",
 		"qwen3vl", "qwen3vlmoe",
 	}, f.KV().String("general.architecture"))
 }

go.mod

@@ -13,7 +13,7 @@ require (
 	github.com/mattn/go-sqlite3 v1.14.24
 	github.com/olekukonko/tablewriter v0.0.5
 	github.com/spf13/cobra v1.7.0
-	github.com/stretchr/testify v1.9.0
+	github.com/stretchr/testify v1.10.0
 	github.com/x448/float16 v0.8.4
 	golang.org/x/sync v0.17.0
 	golang.org/x/sys v0.37.0
@@ -21,13 +21,18 @@ require (
 
 require (
 	github.com/agnivade/levenshtein v1.1.1
+	github.com/charmbracelet/bubbletea v1.3.10
+	github.com/charmbracelet/lipgloss v1.1.0
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
 	github.com/dlclark/regexp2 v1.11.4
 	github.com/emirpasic/gods/v2 v2.0.0-alpha
-	github.com/mattn/go-runewidth v0.0.14
+	github.com/mattn/go-runewidth v0.0.16
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
+	github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c
 	github.com/tkrajina/typescriptify-golang-structs v0.2.0
+	github.com/tree-sitter/go-tree-sitter v0.25.0
+	github.com/tree-sitter/tree-sitter-cpp v0.23.4
 	github.com/wk8/go-ordered-map/v2 v2.1.8
 	golang.org/x/image v0.22.0
 	golang.org/x/mod v0.30.0
@@ -37,22 +42,35 @@ require (
 
 require (
 	github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 // indirect
+	github.com/aymanbagabas/go-osc52/v2 v2.0.1 // indirect
 	github.com/bahlo/generic-list-go v0.2.0 // indirect
 	github.com/buger/jsonparser v1.1.1 // indirect
 	github.com/bytedance/sonic/loader v0.1.1 // indirect
+	github.com/charmbracelet/colorprofile v0.2.3-0.20250311203215-f60798e515dc // indirect
+	github.com/charmbracelet/x/ansi v0.10.1 // indirect
+	github.com/charmbracelet/x/cellbuf v0.0.13-0.20250311204145-2c3ea96c31dd // indirect
+	github.com/charmbracelet/x/term v0.2.1 // indirect
 	github.com/chewxy/hm v1.0.0 // indirect
 	github.com/chewxy/math32 v1.11.0 // indirect
 	github.com/cloudwego/base64x v0.1.4 // indirect
 	github.com/cloudwego/iasm v0.2.0 // indirect
 	github.com/davecgh/go-spew v1.1.1 // indirect
+	github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f // indirect
 	github.com/gogo/protobuf v1.3.2 // indirect
 	github.com/google/flatbuffers v24.3.25+incompatible // indirect
 	github.com/kr/text v0.2.0 // indirect
+	github.com/lucasb-eyer/go-colorful v1.2.0 // indirect
 	github.com/mailru/easyjson v0.7.7 // indirect
+	github.com/mattn/go-localereader v0.0.1 // indirect
+	github.com/mattn/go-pointer v0.0.1 // indirect
+	github.com/muesli/ansi v0.0.0-20230316100256-276c6243b2f6 // indirect
+	github.com/muesli/cancelreader v0.2.2 // indirect
+	github.com/muesli/termenv v0.16.0 // indirect
 	github.com/pkg/errors v0.9.1 // indirect
 	github.com/pmezard/go-difflib v1.0.0 // indirect
-	github.com/rivo/uniseg v0.2.0 // indirect
+	github.com/rivo/uniseg v0.4.7 // indirect
 	github.com/tkrajina/go-reflector v0.5.5 // indirect
+	github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect
 	github.com/xtgo/set v1.0.0 // indirect
 	go4.org/unsafe/assume-no-moving-gc v0.0.0-20231121144256-b99613f794b6 // indirect
 	golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1 // indirect

go.sum

@@ -14,6 +14,8 @@ github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 h1:q4dksr6IC
 github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40/go.mod h1:Q7yQnSMnLvcXlZ8RV+jwz/6y1rQTqbX6C82SndT52Zs=
 github.com/arbovm/levenshtein v0.0.0-20160628152529-48b4e1c0c4d0 h1:jfIu9sQUG6Ig+0+Ap1h4unLjW6YQJpKZVmUzxsD4E/Q=
 github.com/arbovm/levenshtein v0.0.0-20160628152529-48b4e1c0c4d0/go.mod h1:t2tdKJDJF9BV14lnkjHmOQgcvEKgtqs5a1N3LNdJhGE=
+github.com/aymanbagabas/go-osc52/v2 v2.0.1 h1:HwpRHbFMcZLEVr42D4p7XBqjyuxQH5SMiErDT4WkJ2k=
+github.com/aymanbagabas/go-osc52/v2 v2.0.1/go.mod h1:uYgXzlJ7ZpABp8OJ+exZzJJhRNQ2ASbcXHWsFqH8hp8=
 github.com/bahlo/generic-list-go v0.2.0 h1:5sz/EEAK+ls5wF+NeqDpk5+iNdMDXrh3z3nPnH1Wvgk=
 github.com/bahlo/generic-list-go v0.2.0/go.mod h1:2KvAjgMlE5NNynlg/5iLrrCCZ2+5xWbdbCW3pNTGyYg=
 github.com/boombuler/barcode v1.0.0/go.mod h1:paBWMcWSl3LHKBqUq+rly7CNSldXjb2rDl3JlRe0mD8=
@@ -24,6 +26,18 @@ github.com/bytedance/sonic v1.11.6/go.mod h1:LysEHSvpvDySVdC2f87zGWf6CIKJcAvqab1
 github.com/bytedance/sonic/loader v0.1.1 h1:c+e5Pt1k/cy5wMveRDyk2X4B9hF4g7an8N3zCYjJFNM=
 github.com/bytedance/sonic/loader v0.1.1/go.mod h1:ncP89zfokxS5LZrJxl5z0UJcsk4M4yY2JpfqGeCtNLU=
 github.com/census-instrumentation/opencensus-proto v0.2.1/go.mod h1:f6KPmirojxKA12rnyqOA5BBL4O983OfeGPqjHWSTneU=
+github.com/charmbracelet/bubbletea v1.3.10 h1:otUDHWMMzQSB0Pkc87rm691KZ3SWa4KUlvF9nRvCICw=
+github.com/charmbracelet/bubbletea v1.3.10/go.mod h1:ORQfo0fk8U+po9VaNvnV95UPWA1BitP1E0N6xJPlHr4=
+github.com/charmbracelet/colorprofile v0.2.3-0.20250311203215-f60798e515dc h1:4pZI35227imm7yK2bGPcfpFEmuY1gc2YSTShr4iJBfs=
+github.com/charmbracelet/colorprofile v0.2.3-0.20250311203215-f60798e515dc/go.mod h1:X4/0JoqgTIPSFcRA/P6INZzIuyqdFY5rm8tb41s9okk=
+github.com/charmbracelet/lipgloss v1.1.0 h1:vYXsiLHVkK7fp74RkV7b2kq9+zDLoEU4MZoFqR/noCY=
+github.com/charmbracelet/lipgloss v1.1.0/go.mod h1:/6Q8FR2o+kj8rz4Dq0zQc3vYf7X+B0binUUBwA0aL30=
+github.com/charmbracelet/x/ansi v0.10.1 h1:rL3Koar5XvX0pHGfovN03f5cxLbCF2YvLeyz7D2jVDQ=
+github.com/charmbracelet/x/ansi v0.10.1/go.mod h1:3RQDQ6lDnROptfpWuUVIUG64bD2g2BgntdxH0Ya5TeE=
+github.com/charmbracelet/x/cellbuf v0.0.13-0.20250311204145-2c3ea96c31dd h1:vy0GVL4jeHEwG5YOXDmi86oYw2yuYUGqz6a8sLwg0X8=
+github.com/charmbracelet/x/cellbuf v0.0.13-0.20250311204145-2c3ea96c31dd/go.mod h1:xe0nKWGd3eJgtqZRaN9RjMtK7xUYchjzPr7q6kcvCCs=
+github.com/charmbracelet/x/term v0.2.1 h1:AQeHeLZ1OqSXhrAWpYUtZyX1T3zVxfpZuEQMIQaGIAQ=
+github.com/charmbracelet/x/term v0.2.1/go.mod h1:oQ4enTYFV7QN4m0i9mzHrViD7TQKvNEEkHUMCmsxdUg=
 github.com/chewxy/hm v1.0.0 h1:zy/TSv3LV2nD3dwUEQL2VhXeoXbb9QkpmdRAVUFiA6k=
 github.com/chewxy/hm v1.0.0/go.mod h1:qg9YI4q6Fkj/whwHR1D+bOGeF7SniIP40VweVepLjg0=
 github.com/chewxy/math32 v1.0.0/go.mod h1:Miac6hA1ohdDUTagnvJy/q+aNnEk16qWUdb8ZVhvCN0=
@@ -59,6 +73,8 @@ github.com/envoyproxy/go-control-plane v0.9.9-0.20201210154907-fd9021fe5dad/go.m
 github.com/envoyproxy/go-control-plane v0.9.9-0.20210217033140-668b12f5399d/go.mod h1:cXg6YxExXjJnVBQHBLXeUAgxn2UodCpnH306RInaBQk=
 github.com/envoyproxy/go-control-plane v0.9.9-0.20210512163311-63b5d3c536b0/go.mod h1:hliV/p42l8fGbc6Y9bQ70uLwIvmJyVE5k4iMKlh8wCQ=
 github.com/envoyproxy/protoc-gen-validate v0.1.0/go.mod h1:iSmxcyjqTsJpI2R4NaDN7+kN2VEUnK/pcBlmesArF7c=
+github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f h1:Y/CXytFA4m6baUTXGLOoWe4PQhGxaX0KpnayAqC48p4=
+github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f/go.mod h1:vw97MGsxSvLiUE2X8qFplwetxpGLQrlU1Q9AUEIzCaM=
 github.com/fogleman/gg v1.2.1-0.20190220221249-0403632d5b90/go.mod h1:R/bRT+9gY/C5z7JzPU0zXsXHKM4/ayA+zqcVNZzPa1k=
 github.com/fogleman/gg v1.3.0/go.mod h1:R/bRT+9gY/C5z7JzPU0zXsXHKM4/ayA+zqcVNZzPa1k=
 github.com/gabriel-vasile/mimetype v1.4.3 h1:in2uUcidCuFcDKtdcBxlR0rJ1+fsokWf+uqxgUFjbI0=
@@ -148,13 +164,19 @@ github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728 h1:QwWKgMY28TAXaDl+
 github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728/go.mod h1:1fEHWurg7pvf5SG6XNE5Q8UZmOwex51Mkx3SLhrW5B4=
 github.com/leodido/go-urn v1.4.0 h1:WT9HwE9SGECu3lg4d/dIA+jxlljEa1/ffXKmRjqdmIQ=
 github.com/leodido/go-urn v1.4.0/go.mod h1:bvxc+MVxLKB4z00jd1z+Dvzr47oO32F/QSNjSBOlFxI=
+github.com/lucasb-eyer/go-colorful v1.2.0 h1:1nnpGOrhyZZuNyfu1QjKiUICQ74+3FNCN69Aj6K7nkY=
+github.com/lucasb-eyer/go-colorful v1.2.0/go.mod h1:R4dSotOR9KMtayYi1e77YzuveK+i7ruzyGqttikkLy0=
 github.com/mailru/easyjson v0.7.7 h1:UGYAvKxe3sBsEDzO8ZeWOSlIQfWFlxbzLZe7hwFURr0=
 github.com/mailru/easyjson v0.7.7/go.mod h1:xzfreul335JAWq5oZzymOObrkdz5UnU4kGfJJLY9Nlc=
 github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWEY=
 github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
+github.com/mattn/go-localereader v0.0.1 h1:ygSAOl7ZXTx4RdPYinUpg6W99U8jWvWi9Ye2JC/oIi4=
+github.com/mattn/go-localereader v0.0.1/go.mod h1:8fBrzywKY7BI3czFoHkuzRoWE9C+EiG4R1k4Cjx5p88=
+github.com/mattn/go-pointer v0.0.1 h1:n+XhsuGeVO6MEAp7xyEukFINEa+Quek5psIR/ylA6o0=
+github.com/mattn/go-pointer v0.0.1/go.mod h1:2zXcozF6qYGgmsG+SeTZz3oAbFLdD3OWqnUbNvJZAlc=
 github.com/mattn/go-runewidth v0.0.9/go.mod h1:H031xJmbD/WCDINGzjvQ9THkh0rPKHF+m2gUSrubnMI=
-github.com/mattn/go-runewidth v0.0.14 h1:+xnbZSEeDbOIg5/mE6JF0w6n9duR1l3/WmbinWVwUuU=
-github.com/mattn/go-runewidth v0.0.14/go.mod h1:Jdepj2loyihRzMpdS35Xk/zdY8IAYHsh153qUoGf23w=
+github.com/mattn/go-runewidth v0.0.16 h1:E5ScNMtiwvlvB5paMFdw9p4kSQzbXFikJ5SQO6TULQc=
+github.com/mattn/go-runewidth v0.0.16/go.mod h1:Jdepj2loyihRzMpdS35Xk/zdY8IAYHsh153qUoGf23w=
 github.com/mattn/go-sqlite3 v1.14.24 h1:tpSp2G2KyMnnQu99ngJ47EIkWVmliIizyZBfPrBWDRM=
 github.com/mattn/go-sqlite3 v1.14.24/go.mod h1:Uh1q+B4BYcTPb+yiD3kU8Ct7aC0hY9fxUwlHK0RXw+Y=
 github.com/modern-go/concurrent v0.0.0-20180228061459-e0a39a4cb421/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
@@ -162,6 +184,12 @@ github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd h1:TRLaZ9cD/w
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/reflect2 v1.0.2 h1:xBagoLtFs94CBntxluKeaWgTMpvLxC4ur3nMaC9Gz0M=
 github.com/modern-go/reflect2 v1.0.2/go.mod h1:yWuevngMOJpCy52FWWMvUC8ws7m/LJsjYzDa0/r8luk=
+github.com/muesli/ansi v0.0.0-20230316100256-276c6243b2f6 h1:ZK8zHtRHOkbHy6Mmr5D264iyp3TiX5OmNcI5cIARiQI=
+github.com/muesli/ansi v0.0.0-20230316100256-276c6243b2f6/go.mod h1:CJlz5H+gyd6CUWT45Oy4q24RdLyn7Md9Vj2/ldJBSIo=
+github.com/muesli/cancelreader v0.2.2 h1:3I4Kt4BQjOR54NavqnDogx/MIoWBFa0StPA8ELUXHmA=
+github.com/muesli/cancelreader v0.2.2/go.mod h1:3XuTXfFS2VjM+HTLZY9Ak0l6eUKfijIfMUZ4EgX0QYo=
+github.com/muesli/termenv v0.16.0 h1:S5AlUN9dENB57rsbnkPyfdGuWIlkmzJjbFf0Tf5FWUc=
+github.com/muesli/termenv v0.16.0/go.mod h1:ZRfOIKPFDYQoDFF4Olj7/QJbW60Ol/kL1pU3VfY/Cnk=
 github.com/nlpodyssey/gopickle v0.3.0 h1:BLUE5gxFLyyNOPzlXxt6GoHEMMxD0qhsE4p0CIQyoLw=
 github.com/nlpodyssey/gopickle v0.3.0/go.mod h1:f070HJ/yR+eLi5WmM1OXJEGaTpuJEUiib19olXgYha0=
 github.com/olekukonko/tablewriter v0.0.5 h1:P2Ga83D34wi1o9J6Wh1mRuqd4mF/x/lgBS7N7AbDhec=
@@ -174,14 +202,17 @@ github.com/phpdave11/gofpdf v1.4.2/go.mod h1:zpO6xFn9yxo3YLyMvW8HcKWVdbNqgIfOOp2
 github.com/phpdave11/gofpdi v1.0.12/go.mod h1:vBmVV0Do6hSBHC8uKUQ71JGW+ZGQq74llk/7bXwjDoI=
 github.com/pierrec/lz4/v4 v4.1.8 h1:ieHkV+i2BRzngO4Wd/3HGowuZStgq6QkPsD1eolNAO4=
 github.com/pierrec/lz4/v4 v4.1.8/go.mod h1:gZWDp/Ze/IJXGXf23ltt2EXimqmTUXEy0GFuRQyBid4=
+github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c h1:+mdjkGKdHQG3305AYmdv1U2eRNDiU2ErMBj1gwrq8eQ=
+github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c/go.mod h1:7rwL4CYBLnjLxUqIJNnCWiEdr3bn6IUYi15bNlnbCCU=
 github.com/pkg/errors v0.8.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=
 github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/prometheus/client_model v0.0.0-20190812154241-14fe0d1b01d4/go.mod h1:xMI15A0UPsDsEKsMN9yxemIoYk6Tm2C1GtYGdfGttqA=
-github.com/rivo/uniseg v0.2.0 h1:S1pD9weZBuJdFmowNwbpi7BJ8TNftyUImj/0WQi72jY=
 github.com/rivo/uniseg v0.2.0/go.mod h1:J6wj4VEh+S6ZtnVlnTBMWIodfgj8LQOQFoIToxlJtxc=
+github.com/rivo/uniseg v0.4.7 h1:WUdvkW8uEhrYfLC4ZzdpI2ztxP1I582+49Oc5Mq64VQ=
+github.com/rivo/uniseg v0.4.7/go.mod h1:FN3SvrM+Zdj16jyLfmOkMNblXMcoc8DfTHruCPUcx88=
 github.com/rogpeppe/fastuuid v1.2.0/go.mod h1:jVj6XXZzXRy/MSR5jhDC/2q6DgLz+nrA6LYCDYWNEvQ=
 github.com/rogpeppe/go-internal v1.8.0 h1:FCbCCtXNOY3UtUuHUYaghJg4y7Fd14rXifAYUAtL9R8=
 github.com/rogpeppe/go-internal v1.8.0/go.mod h1:WmiCO8CzOY8rg0OYDC4/i/2WRWAB6poM+XZ2dLUbcbE=
@@ -204,12 +235,39 @@ github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/
 github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
 github.com/stretchr/testify v1.8.1/go.mod h1:w2LPCIKwWwSfY2zedu0+kehJoqGctiVI29o6fzry7u4=
 github.com/stretchr/testify v1.8.4/go.mod h1:sz/lmYIOXD/1dqDmKjjqLyZ2RngseejIcXlSw2iwfAo=
-github.com/stretchr/testify v1.9.0 h1:HtqpIVDClZ4nwg75+f6Lvsy/wHu+3BoSGCbBAcpTsTg=
 github.com/stretchr/testify v1.9.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
+github.com/stretchr/testify v1.10.0 h1:Xv5erBjTwe/5IxqUQTdXv5kgmIvbHo3QQyRwhJsOfJA=
+github.com/stretchr/testify v1.10.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
 github.com/tkrajina/go-reflector v0.5.5 h1:gwoQFNye30Kk7NrExj8zm3zFtrGPqOkzFMLuQZg1DtQ=
 github.com/tkrajina/go-reflector v0.5.5/go.mod h1:ECbqLgccecY5kPmPmXg1MrHW585yMcDkVl6IvJe64T4=
 github.com/tkrajina/typescriptify-golang-structs v0.2.0 h1:ZedWk82egydDspGTryAatbX0/1NZDQbdiZLoCbOk4f8=
 github.com/tkrajina/typescriptify-golang-structs v0.2.0/go.mod h1:sjU00nti/PMEOZb07KljFlR+lJ+RotsC0GBQMv9EKls=
+github.com/tree-sitter/go-tree-sitter v0.25.0 h1:sx6kcg8raRFCvc9BnXglke6axya12krCJF5xJ2sftRU=
+github.com/tree-sitter/go-tree-sitter v0.25.0/go.mod h1:r77ig7BikoZhHrrsjAnv8RqGti5rtSyvDHPzgTPsUuU=
+github.com/tree-sitter/tree-sitter-c v0.23.4 h1:nBPH3FV07DzAD7p0GfNvXM+Y7pNIoPenQWBpvM++t4c=
+github.com/tree-sitter/tree-sitter-c v0.23.4/go.mod h1:MkI5dOiIpeN94LNjeCp8ljXN/953JCwAby4bClMr6bw=
+github.com/tree-sitter/tree-sitter-cpp v0.23.4 h1:LaWZsiqQKvR65yHgKmnaqA+uz6tlDJTJFCyFIeZU/8w=
+github.com/tree-sitter/tree-sitter-cpp v0.23.4/go.mod h1:doqNW64BriC7WBCQ1klf0KmJpdEvfxyXtoEybnBo6v8=
+github.com/tree-sitter/tree-sitter-embedded-template v0.23.2 h1:nFkkH6Sbe56EXLmZBqHHcamTpmz3TId97I16EnGy4rg=
+github.com/tree-sitter/tree-sitter-embedded-template v0.23.2/go.mod h1:HNPOhN0qF3hWluYLdxWs5WbzP/iE4aaRVPMsdxuzIaQ=
+github.com/tree-sitter/tree-sitter-go v0.23.4 h1:yt5KMGnTHS+86pJmLIAZMWxukr8W7Ae1STPvQUuNROA=
+github.com/tree-sitter/tree-sitter-go v0.23.4/go.mod h1:Jrx8QqYN0v7npv1fJRH1AznddllYiCMUChtVjxPK040=
+github.com/tree-sitter/tree-sitter-html v0.23.2 h1:1UYDV+Yd05GGRhVnTcbP58GkKLSHHZwVaN+lBZV11Lc=
+github.com/tree-sitter/tree-sitter-html v0.23.2/go.mod h1:gpUv/dG3Xl/eebqgeYeFMt+JLOY9cgFinb/Nw08a9og=
+github.com/tree-sitter/tree-sitter-java v0.23.5 h1:J9YeMGMwXYlKSP3K4Us8CitC6hjtMjqpeOf2GGo6tig=
+github.com/tree-sitter/tree-sitter-java v0.23.5/go.mod h1:NRKlI8+EznxA7t1Yt3xtraPk1Wzqh3GAIC46wxvc320=
+github.com/tree-sitter/tree-sitter-javascript v0.23.1 h1:1fWupaRC0ArlHJ/QJzsfQ3Ibyopw7ZfQK4xXc40Zveo=
+github.com/tree-sitter/tree-sitter-javascript v0.23.1/go.mod h1:lmGD1EJdCA+v0S1u2fFgepMg/opzSg/4pgFym2FPGAs=
+github.com/tree-sitter/tree-sitter-json v0.24.8 h1:tV5rMkihgtiOe14a9LHfDY5kzTl5GNUYe6carZBn0fQ=
+github.com/tree-sitter/tree-sitter-json v0.24.8/go.mod h1:F351KK0KGvCaYbZ5zxwx/gWWvZhIDl0eMtn+1r+gQbo=
+github.com/tree-sitter/tree-sitter-php v0.23.11 h1:iHewsLNDmznh8kgGyfWfujsZxIz1YGbSd2ZTEM0ZiP8=
+github.com/tree-sitter/tree-sitter-php v0.23.11/go.mod h1:T/kbfi+UcCywQfUNAJnGTN/fMSUjnwPXA8k4yoIks74=
+github.com/tree-sitter/tree-sitter-python v0.23.6 h1:qHnWFR5WhtMQpxBZRwiaU5Hk/29vGju6CVtmvu5Haas=
+github.com/tree-sitter/tree-sitter-python v0.23.6/go.mod h1:cpdthSy/Yoa28aJFBscFHlGiU+cnSiSh1kuDVtI8YeM=
+github.com/tree-sitter/tree-sitter-ruby v0.23.1 h1:T/NKHUA+iVbHM440hFx+lzVOzS4dV6z8Qw8ai+72bYo=
+github.com/tree-sitter/tree-sitter-ruby v0.23.1/go.mod h1:kUS4kCCQloFcdX6sdpr8p6r2rogbM6ZjTox5ZOQy8cA=
+github.com/tree-sitter/tree-sitter-rust v0.23.2 h1:6AtoooCW5GqNrRpfnvl0iUhxTAZEovEmLKDbyHlfw90=
+github.com/tree-sitter/tree-sitter-rust v0.23.2/go.mod h1:hfeGWic9BAfgTrc7Xf6FaOAguCFJRo3RBbs7QJ6D7MI=
 github.com/twitchyliquid64/golang-asm v0.15.1 h1:SU5vSMR7hnwNxj24w34ZyCi/FmDZTkS4MhqMhdFk5YI=
 github.com/twitchyliquid64/golang-asm v0.15.1/go.mod h1:a1lVb/DtPvCB8fslRZhAngC2+aY1QWCk3Cedj/Gdt08=
 github.com/ugorji/go/codec v1.2.12 h1:9LC83zGrHhuUA9l16C9AHXAqEV/2wBQ4nkvumAE65EE=
@@ -218,6 +276,8 @@ github.com/wk8/go-ordered-map/v2 v2.1.8 h1:5h/BUHu93oj4gIdvHHHGsScSTMijfx5PeYkE/
 github.com/wk8/go-ordered-map/v2 v2.1.8/go.mod h1:5nJHM5DyteebpVlHnWMV0rPz6Zp7+xBAnxjb1X5vnTw=
 github.com/x448/float16 v0.8.4 h1:qLwI1I70+NjRFUR3zs1JPUCgaCXSh3SW62uAKT1mSBM=
 github.com/x448/float16 v0.8.4/go.mod h1:14CWIYCyZA/cWjXOioeEpHeN/83MdbZDRQHoFcYsOfg=
+github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e h1:JVG44RsyaB9T2KIHavMF/ppJZNG9ZpyihvCd0w101no=
+github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e/go.mod h1:RbqR21r5mrJuqunuUZ/Dhy/avygyECGrLceyNeo4LiM=
 github.com/xtgo/set v1.0.0 h1:6BCNBRv3ORNDQ7fyoJXRv+tstJz3m1JVFQErfeZz2pY=
 github.com/xtgo/set v1.0.0/go.mod h1:d3NHzGzSa0NmB2NhFyECA+QdRp29oEn2xbT+TpeFoM8=
 github.com/yuin/goldmark v1.1.27/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74=
@@ -304,6 +364,8 @@ golang.org/x/sys v0.0.0-20210330210617-4fbd30eecc44/go.mod h1:h1NjWce9XRLGQEsW7w
 golang.org/x/sys v0.0.0-20210423082822-04245dca01da/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20210510120138-977fb7262007/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.0.0-20210809222454-d867a43fc93e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.1.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.37.0 h1:fdNQudmxPjkdUTPnLn5mdQv7Zwvbvpaxqs831goi9kQ=

integration/basic_test.go

@@ -144,3 +144,47 @@ func TestUnicodeModelDir(t *testing.T) {
 	}
 	ChatTestHelper(ctx, t, req, blueSkyExpected)
 }
+
+// TestNumPredict verifies that when num_predict is set, the model generates
+// exactly that many tokens. It uses logprobs to count the actual tokens output.
+func TestNumPredict(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	if err := PullIfMissing(ctx, client, "qwen3:0.6b"); err != nil {
+		t.Fatalf("failed to pull model: %v", err)
+	}
+
+	req := api.GenerateRequest{
+		Model:    "qwen3:0.6b",
+		Prompt:   "Write a long story.",
+		Stream:   &stream,
+		Logprobs: true,
+		Options: map[string]any{
+			"num_predict": 10,
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+
+	logprobCount := 0
+	var finalResponse api.GenerateResponse
+	err := client.Generate(ctx, &req, func(resp api.GenerateResponse) error {
+		logprobCount += len(resp.Logprobs)
+		if resp.Done {
+			finalResponse = resp
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("generate failed: %v", err)
+	}
+
+	if logprobCount != 10 {
+		t.Errorf("expected 10 tokens (logprobs), got %d (EvalCount=%d, DoneReason=%s)",
+			logprobCount, finalResponse.EvalCount, finalResponse.DoneReason)
+	}
+}

integration/imagegen_test.go

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

integration/utils_test.go

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

kvcache/cache.go

@@ -75,3 +75,10 @@ type Cache interface {
 	// removed by calling Remove(seq, 0, math.MaxInt32)
 	Remove(seq int, beginIndex, endIndex int32) error
 }
+
+// CheckpointCache optionally supports restoring recurrent state to a prior
+// position to avoid full prompt reprocessing when a prefix mismatch occurs.
+// The returned position is the number of tokens that can be kept (prefix length).
+type CheckpointCache interface {
+	PrepareRestore(seq int, targetPos int32) (int32, bool)
+}

llama/patches/0032-ggml-enable-MLA-flash-attention-for-GLM-4.7-flash.patch

@@ -0,0 +1,309 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: nobody <>
+Date: Sat, 24 Jan 2026 02:31:01 +0000
+Subject: [PATCH] ggml: enable MLA flash attention for GLM-4.7-flash
+
+Add support for gqa_ratio 4 in MLA flash attention kernels. GLM-4.7-flash
+uses head size 576 with gqa_ratio 4, which was previously only supported
+for gqa_ratio 16 (DeepSeek).
+
+Metal changes:
+- Enable head size 576 for flash attention
+- Increase simdgroups to 8 for large heads (>=512)
+- Add case 8 kernel dispatch for 8 simdgroups
+
+CUDA changes:
+- Add gqa_ratio 4 support for head 576/512
+- Add tile configs for (576, 512, 4) and (576, 512, 8)
+- Add MMA config cases for ncols 4
+- Add template instances for ncols2=4
+- Fix nbatch_fa values in nvidia_fp32 config (32->64)
+---
+ ggml/src/ggml-cuda/fattn-mma-f16.cuh          | 40 +++++++++++++++----
+ ggml/src/ggml-cuda/fattn-tile.cuh             | 16 ++++++++
+ ggml/src/ggml-cuda/fattn.cu                   | 12 ++++--
+ ...ttn-mma-f16-instance-ncols1_16-ncols2_4.cu |  1 +
+ ...attn-mma-f16-instance-ncols1_2-ncols2_4.cu |  1 +
+ ...attn-mma-f16-instance-ncols1_4-ncols2_4.cu |  1 +
+ ...attn-mma-f16-instance-ncols1_8-ncols2_4.cu |  1 +
+ ggml/src/ggml-metal/ggml-metal-device.m       |  8 +---
+ ggml/src/ggml-metal/ggml-metal-ops.cpp        |  2 +-
+ ggml/src/ggml-metal/ggml-metal.metal          |  1 +
+ 10 files changed, 64 insertions(+), 19 deletions(-)
+
+diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+index 7bd1044c1..3dea2205e 100644
+--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
++++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+@@ -66,7 +66,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  32, 128, 128, 128, 2, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  32, 128, 128, 128, 2, true);
+ 
+-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
+@@ -80,7 +81,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
+ 
+-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32,  96,  64, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
+@@ -89,7 +91,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
+ }
+ 
+ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_volta(const int DKQ, const int DV, const int ncols) {
+-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256,  64, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256,  64, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128,  64, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128,  64, 1, false);
+@@ -397,7 +400,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
+     constexpr int  ncols           = ncols1 * ncols2;
+     constexpr int  cols_per_warp   = T_B_KQ::I;
+     constexpr int  cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
+-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
++    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
+     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
+     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2(DKQ, DV, ncols);
+     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2(DKQ, DV, ncols);
+@@ -467,7 +470,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
+                 }
+             }
+         } else {
+-            static_assert(cols_per_warp != 8, "cols_per_warp == 8 not implemented");
+ #pragma unroll
+             for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) {
+                 load_ldmatrix(Q_B[0], tile_Q + (threadIdx.y / np)*(T_B_KQ::I*stride_tile_Q) + k_KQ_0, stride_tile_Q);
+@@ -479,8 +481,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
+                     T_A_KQ K_A;
+                     load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
+ 
+-                    // Wide version of KQ_C is column-major => swap A and B.
+-                    mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
++                    if constexpr (cols_per_warp == 8) {
++                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
++                    } else {
++                        // Wide version of KQ_C is column-major
++#if defined(AMD_WMMA_AVAILABLE)
++                        // RDNA matrix C is column-major.
++                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
++#else
++                        // swap A and B for CUDA.
++                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
++#endif // defined(AMD_WMMA_AVAILABLE)
++                    }
+                 }
+             }
+         }
+@@ -841,7 +853,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
+ 
+     constexpr int  cols_per_warp   = T_B_KQ::I;
+     constexpr int  cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
+-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
++    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
+     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa     (DKQ, DV, ncols);
+     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2     (DKQ, DV, ncols);
+     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2     (DKQ, DV, ncols);
+@@ -1353,6 +1365,13 @@ static __global__ void flash_attn_ext_f16(
+         NO_DEVICE_CODE;
+         return;
+     }
++#ifdef VOLTA_MMA_AVAILABLE
++    if (ncols1*ncols2 < 32) {
++        NO_DEVICE_CODE;
++        return;
++    }
++#endif // VOLTA_MMA_AVAILABLE
++
+ #if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+     if (ncols1*ncols2 > 32) {
+         NO_DEVICE_CODE;
+@@ -1585,3 +1604,8 @@ DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256,  64)
+ extern DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);
+ extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);
+ extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);
++
++// For GLM 4.7 Flash
++extern DECL_FATTN_MMA_F16_CASE(576, 512,  4,  4);
++extern DECL_FATTN_MMA_F16_CASE(576, 512,  8,  4);
++extern DECL_FATTN_MMA_F16_CASE(576, 512, 16,  4);
+diff --git a/ggml/src/ggml-cuda/fattn-tile.cuh b/ggml/src/ggml-cuda/fattn-tile.cuh
+index 7c4d6fe67..371be7442 100644
+--- a/ggml/src/ggml-cuda/fattn-tile.cuh
++++ b/ggml/src/ggml-cuda/fattn-tile.cuh
+@@ -68,6 +68,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  64,  64)
+ 
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
+ 
+     return 0;
+@@ -122,6 +124,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32,  64)
+ 
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  32,  64)
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  32,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  32,  64)
+ 
+     return 0;
+@@ -183,6 +187,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
+ 
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128,  64)
+ 
+@@ -245,6 +251,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
+ 
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
++    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4,  64,  64)
+     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128,  64)
+ 
+@@ -1187,6 +1195,14 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
+             launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
+             return;
+         }
++        if (use_gqa_opt && gqa_ratio % 8 == 0) {
++            launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
++            return;
++        }
++        if (use_gqa_opt && gqa_ratio % 4 == 0) {
++            launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
++            return;
++        }
+     }
+ 
+     if constexpr (DV <= 256) {
+diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
+index 015540666..1693479cb 100644
+--- a/ggml/src/ggml-cuda/fattn.cu
++++ b/ggml/src/ggml-cuda/fattn.cu
+@@ -111,7 +111,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
+             ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
+             break;
+         case 576: {
+-            // For Deepseek, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
++            // For Deepseek/GLM4, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
+             GGML_ASSERT(V->ne[0] == 512);
+             float max_bias = 0.0f;
+             memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+@@ -121,8 +121,12 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
+ 
+             GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
+             const int gqa_ratio = Q->ne[2] / K->ne[2];
+-            GGML_ASSERT(gqa_ratio % 16 == 0);
+-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
++            GGML_ASSERT(gqa_ratio % 4 == 0);
++            if (gqa_ratio % 16 == 0) {
++                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
++            } else {
++                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512,  4>(ctx, dst);
++            }
+         } break;
+         default:
+             GGML_ABORT("fatal error");
+@@ -251,7 +255,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
+             if (V->ne[0] != 512) {
+                 return BEST_FATTN_KERNEL_NONE;
+             }
+-            if (!gqa_opt_applies || gqa_ratio % 16 != 0) {
++            if (!gqa_opt_applies || gqa_ratio % 4 != 0) {
+                 return BEST_FATTN_KERNEL_NONE;
+             }
+             break;
+diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
+index 2074e954a..517993cb0 100644
+--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
++++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
+@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
+ DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
+ DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
+ DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);
++DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4);
+diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
+index 24c64cf00..97b19c67a 100644
+--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
++++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
+@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
+ DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
+ DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
+ DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);
++DECL_FATTN_MMA_F16_CASE(576, 512, 2, 4);
+diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
+index 1ada657f1..989626dfa 100644
+--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
++++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
+@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
+ DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
+ DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
+ DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);
++DECL_FATTN_MMA_F16_CASE(576, 512, 4, 4);
+diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
+index 86d4ffae2..173de7aac 100644
+--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
++++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
+@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 8, 4);
+ DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4);
+ DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4);
+ DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4);
++DECL_FATTN_MMA_F16_CASE(576, 512, 8, 4);
+diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m
+index f24270bb1..7b5ee968c 100644
+--- a/ggml/src/ggml-metal/ggml-metal-device.m
++++ b/ggml/src/ggml-metal/ggml-metal-device.m
+@@ -1071,12 +1071,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
+                 op->src[0]->ne[0] != 112 &&
+                 op->src[0]->ne[0] != 128 &&
+                 op->src[0]->ne[0] != 192 &&
+-                op->src[0]->ne[0] != 256) {
+-                return false;
+-            }
+-            if (op->src[0]->ne[0] == 576) {
+-                // DeepSeek sizes
+-                // TODO: disabled for now, until optmized
++                op->src[0]->ne[0] != 256 &&
++                op->src[0]->ne[0] != 576) {
+                 return false;
+             }
+             if (op->src[1]->type != op->src[2]->type) {
+diff --git a/ggml/src/ggml-metal/ggml-metal-ops.cpp b/ggml/src/ggml-metal/ggml-metal-ops.cpp
+index e99c1763f..80864f303 100644
+--- a/ggml/src/ggml-metal/ggml-metal-ops.cpp
++++ b/ggml/src/ggml-metal/ggml-metal-ops.cpp
+@@ -2456,7 +2456,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
+ 
+         // simdgroups per threadgroup (a.k.a. warps)
+         //nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
+-        int32_t nsg = 4;
++        int32_t nsg = ne00 >= 512 ? 8 : 4;
+ 
+         const size_t smem = FATTN_SMEM(nsg);
+ 
+diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
+index c98d269d1..d33c16079 100644
+--- a/ggml/src/ggml-metal/ggml-metal.metal
++++ b/ggml/src/ggml-metal/ggml-metal.metal
+@@ -6166,6 +6166,7 @@ kernel void kernel_flash_attn_ext(
+       //case 1: kernel_flash_attn_ext_impl<FWD_TMPL, 1>(FWD_ARGS); break;
+       //case 2: kernel_flash_attn_ext_impl<FWD_TMPL, 2>(FWD_ARGS); break;
+         case 4: kernel_flash_attn_ext_impl<FWD_TMPL, 4>(FWD_ARGS); break;
++        case 8: kernel_flash_attn_ext_impl<FWD_TMPL, 8>(FWD_ARGS); break;
+     }
+ #undef FWD_TMPL
+ #undef FWD_ARGS

llama/patches/0033-ggml-metal-solve_tri.patch

@@ -0,0 +1,276 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Jeffrey Morgan <jmorganca@gmail.com>
+Date: Tue, 3 Feb 2026 12:00:00 -0800
+Subject: [PATCH] ggml: metal solve_tri
+
+---
+ ggml/src/ggml-metal/ggml-metal-device.cpp | 20 +++++++
+ ggml/src/ggml-metal/ggml-metal-device.h   |  1 +
+ ggml/src/ggml-metal/ggml-metal-device.m   | 11 ++++
+ ggml/src/ggml-metal/ggml-metal-impl.h     | 21 ++++++++
+ ggml/src/ggml-metal/ggml-metal-ops.cpp    | 63 +++++++++++++++++++++++
+ ggml/src/ggml-metal/ggml-metal-ops.h      |  1 +
+ ggml/src/ggml-metal/ggml-metal.metal      | 60 +++++++++++++++++++++
+ 7 files changed, 177 insertions(+)
+
+diff --git a/ggml/src/ggml-metal/ggml-metal-device.cpp b/ggml/src/ggml-metal/ggml-metal-device.cpp
+index 680904d13..83385c9ef 100644
+--- a/ggml/src/ggml-metal/ggml-metal-device.cpp
++++ b/ggml/src/ggml-metal/ggml-metal-device.cpp
+@@ -1370,6 +1370,26 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm(ggml_met
+     return res;
+ }
+ 
++ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri(ggml_metal_library_t lib, const ggml_tensor * op) {
++    assert(op->op == GGML_OP_SOLVE_TRI);
++
++    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
++    GGML_ASSERT(ggml_is_contiguous(op->src[1]));
++
++    char base[256];
++    char name[256];
++
++    snprintf(base, 256, "kernel_solve_tri_f32");
++    snprintf(name, 256, "%s", base);
++
++    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
++    if (!res.pipeline) {
++        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
++    }
++
++    return res;
++}
++
+ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_group_norm(ggml_metal_library_t lib, const ggml_tensor * op) {
+     assert(op->op == GGML_OP_GROUP_NORM);
+ 
+diff --git a/ggml/src/ggml-metal/ggml-metal-device.h b/ggml/src/ggml-metal/ggml-metal-device.h
+index 0a8b9211a..8a9d17460 100644
+--- a/ggml/src/ggml-metal/ggml-metal-device.h
++++ b/ggml/src/ggml-metal/ggml-metal-device.h
+@@ -133,6 +133,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k
+ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
+ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
++struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri         (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_norm              (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse);
+ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rope              (ggml_metal_library_t lib, const struct ggml_tensor * op);
+diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m
+index 7b5ee968c..4e5acfbe5 100644
+--- a/ggml/src/ggml-metal/ggml-metal-device.m
++++ b/ggml/src/ggml-metal/ggml-metal-device.m
+@@ -1023,6 +1023,17 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
+             return has_simdgroup_reduction && ggml_is_contiguous_rows(op->src[0]);
+         case GGML_OP_L2_NORM:
+             return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(op->src[0]));
++        case GGML_OP_SOLVE_TRI:
++            return ggml_is_contiguous(op->src[0]) &&
++                ggml_is_contiguous(op->src[1]) &&
++                op->src[0]->type == GGML_TYPE_F32 &&
++                op->src[1]->type == GGML_TYPE_F32 &&
++                op->type == GGML_TYPE_F32;
++        case GGML_OP_COUNT_EQUAL:
++            return has_simdgroup_reduction &&
++                op->src[0]->type == GGML_TYPE_I32 &&
++                op->src[1]->type == GGML_TYPE_I32 &&
++                op->type == GGML_TYPE_I64;
+         case GGML_OP_ARGMAX:
+             return has_simdgroup_reduction;
+         case GGML_OP_NORM:
+diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h
+index 8944b07e9..cfdea9c07 100644
+--- a/ggml/src/ggml-metal/ggml-metal-impl.h
++++ b/ggml/src/ggml-metal/ggml-metal-impl.h
+@@ -500,6 +500,27 @@ typedef struct {
+     float    eps;
+ } ggml_metal_kargs_l2_norm;
+ 
++typedef struct {
++    int32_t  ne00;
++    int32_t  ne01;
++    int32_t  ne02;
++    int32_t  ne03;
++    uint64_t nb00;
++    uint64_t nb01;
++    uint64_t nb02;
++    uint64_t nb03;
++    int32_t  ne10;
++    int32_t  ne11;
++    uint64_t nb10;
++    uint64_t nb11;
++    uint64_t nb12;
++    uint64_t nb13;
++    uint64_t nb0;
++    uint64_t nb1;
++    uint64_t nb2;
++    uint64_t nb3;
++} ggml_metal_kargs_solve_tri;
++
+ typedef struct {
+     int64_t  ne00;
+     int64_t  ne01;
+diff --git a/ggml/src/ggml-metal/ggml-metal-ops.cpp b/ggml/src/ggml-metal/ggml-metal-ops.cpp
+index 80864f303..4ac135603 100644
+--- a/ggml/src/ggml-metal/ggml-metal-ops.cpp
++++ b/ggml/src/ggml-metal/ggml-metal-ops.cpp
+@@ -357,6 +357,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
+             {
+                 n_fuse = ggml_metal_op_l2_norm(ctx, idx);
+             } break;
++        case GGML_OP_SOLVE_TRI:
++            {
++                n_fuse = ggml_metal_op_solve_tri(ctx, idx);
++            } break;
+         case GGML_OP_GROUP_NORM:
+             {
+                 n_fuse = ggml_metal_op_group_norm(ctx, idx);
+@@ -2931,6 +2935,65 @@ int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
+     return 1;
+ }
+ 
++int ggml_metal_op_solve_tri(ggml_metal_op_t ctx, int idx) {
++    ggml_tensor * op = ctx->node(idx);
++
++    ggml_metal_library_t lib = ctx->lib;
++    ggml_metal_encoder_t enc = ctx->enc;
++
++    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
++    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
++    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
++    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
++    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
++    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
++
++    ggml_metal_kargs_solve_tri args = {
++        /*.ne00 =*/ ne00,
++        /*.ne01 =*/ ne01,
++        /*.ne02 =*/ ne02,
++        /*.ne03 =*/ ne03,
++        /*.nb00 =*/ nb00,
++        /*.nb01 =*/ nb01,
++        /*.nb02 =*/ nb02,
++        /*.nb03 =*/ nb03,
++        /*.ne10 =*/ ne10,
++        /*.ne11 =*/ ne11,
++        /*.nb10 =*/ nb10,
++        /*.nb11 =*/ nb11,
++        /*.nb12 =*/ nb12,
++        /*.nb13 =*/ nb13,
++        /*.nb0  =*/ nb0,
++        /*.nb1  =*/ nb1,
++        /*.nb2  =*/ nb2,
++        /*.nb3  =*/ nb3,
++    };
++
++    auto pipeline = ggml_metal_library_get_pipeline_solve_tri(lib, op);
++
++    const int64_t ncols = ne10;
++    const int64_t n_batches = (int64_t)ne02 * ne03;
++    const int64_t nr = n_batches * ncols;
++
++    int nth = 64;
++    nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
++    if (nth < 1) {
++        nth = 1;
++    }
++
++    const int64_t n_tg = (nr + nth - 1) / nth;
++
++    ggml_metal_encoder_set_pipeline(enc, pipeline);
++    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
++    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
++    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
++    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
++
++    ggml_metal_encoder_dispatch_threadgroups(enc, n_tg, 1, 1, nth, 1, 1);
++
++    return 1;
++}
++
+ int ggml_metal_op_group_norm(ggml_metal_op_t ctx, int idx) {
+     ggml_tensor * op = ctx->node(idx);
+ 
+diff --git a/ggml/src/ggml-metal/ggml-metal-ops.h b/ggml/src/ggml-metal/ggml-metal-ops.h
+index 902b54452..a475183d3 100644
+--- a/ggml/src/ggml-metal/ggml-metal-ops.h
++++ b/ggml/src/ggml-metal/ggml-metal-ops.h
+@@ -68,6 +68,7 @@ int ggml_metal_op_add_id            (ggml_metal_op_t ctx, int idx);
+ int ggml_metal_op_flash_attn_ext    (ggml_metal_op_t ctx, int idx);
+ int ggml_metal_op_bin               (ggml_metal_op_t ctx, int idx);
+ int ggml_metal_op_l2_norm           (ggml_metal_op_t ctx, int idx);
++int ggml_metal_op_solve_tri         (ggml_metal_op_t ctx, int idx);
+ int ggml_metal_op_group_norm        (ggml_metal_op_t ctx, int idx);
+ int ggml_metal_op_norm              (ggml_metal_op_t ctx, int idx);
+ int ggml_metal_op_rope              (ggml_metal_op_t ctx, int idx);
+diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
+index d33c16079..c37447a10 100644
+--- a/ggml/src/ggml-metal/ggml-metal.metal
++++ b/ggml/src/ggml-metal/ggml-metal.metal
+@@ -3012,6 +3012,66 @@ kernel void kernel_l2_norm_f32(
+     }
+ }
+ 
++kernel void kernel_solve_tri_f32(
++        constant ggml_metal_kargs_solve_tri & args,
++        device const char * src0,
++        device const char * src1,
++        device       char * dst,
++        uint   tgpig[[threadgroup_position_in_grid]],
++        ushort tpitg[[thread_position_in_threadgroup]],
++        ushort   ntg[[threads_per_threadgroup]]) {
++    const uint64_t ncols = (uint64_t) args.ne10;
++    const uint64_t n_batches = (uint64_t) args.ne02 * (uint64_t) args.ne03;
++    const uint64_t nr = n_batches * ncols;
++
++    const uint64_t gid = (uint64_t) tgpig * (uint64_t) ntg + (uint64_t) tpitg;
++    if (gid >= nr) {
++        return;
++    }
++
++    const uint64_t i03 = gid / ((uint64_t) args.ne02 * ncols);
++    const uint64_t rem = gid - i03 * (uint64_t) args.ne02 * ncols;
++    const uint64_t i02 = rem / ncols;
++    const uint64_t i01 = rem - i02 * ncols;
++
++    const uint64_t sa0 = args.nb00 / sizeof(float);
++    const uint64_t sa1 = args.nb01 / sizeof(float);
++    const uint64_t sa2 = args.nb02 / sizeof(float);
++    const uint64_t sa3 = args.nb03 / sizeof(float);
++
++    const uint64_t sb0 = args.nb10 / sizeof(float);
++    const uint64_t sb1 = args.nb11 / sizeof(float);
++    const uint64_t sb2 = args.nb12 / sizeof(float);
++    const uint64_t sb3 = args.nb13 / sizeof(float);
++
++    const uint64_t sx0 = args.nb0 / sizeof(float);
++    const uint64_t sx1 = args.nb1 / sizeof(float);
++    const uint64_t sx2 = args.nb2 / sizeof(float);
++    const uint64_t sx3 = args.nb3 / sizeof(float);
++
++    device const float * A = (device const float *) src0;
++    device const float * B = (device const float *) src1;
++    device       float * X = (device       float *) dst;
++
++    const uint64_t A_base = i02 * sa2 + i03 * sa3;
++    const uint64_t B_base = i02 * sb2 + i03 * sb3;
++    const uint64_t X_base = i02 * sx2 + i03 * sx3;
++
++    const uint64_t n = (uint64_t) args.ne11;
++
++    for (uint64_t i00 = 0; i00 < n; ++i00) {
++        float sum = 0.0f;
++        for (uint64_t t = 0; t < i00; ++t) {
++            sum += A[A_base + i00 * sa1 + t * sa0] *
++                X[X_base + t * sx1 + i01 * sx0];
++        }
++
++        const float diag = A[A_base + i00 * sa1 + i00 * sa0];
++        X[X_base + i00 * sx1 + i01 * sx0] =
++            (B[B_base + i00 * sb1 + i01 * sb0] - sum) / diag;
++    }
++}
++
+ kernel void kernel_group_norm_f32(
+         constant ggml_metal_kargs_group_norm & args,
+         device const float * src0,

llm/server.go

@@ -34,6 +34,7 @@ import (
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/tokenizer"
 )
 
 type filteredEnv []string
@@ -80,6 +81,7 @@ type LlamaServer interface {
 	GetPort() int
 	GetDeviceInfos(ctx context.Context) []ml.DeviceInfo
 	HasExited() bool
+	ContextLength() int
 }
 
 // llmServer is an instance of a runner hosting a single model
@@ -115,7 +117,7 @@ type llamaServer struct {
 type ollamaServer struct {
 	llmServer
 
-	textProcessor model.TextProcessor // textProcessor handles text encoding/decoding
+	tokenizer tokenizer.Tokenizer // tokenizer handles text encoding/decoding
 }
 
 // LoadModel will load a model from disk. The model must be in the GGML format.
@@ -141,11 +143,11 @@ func LoadModel(model string, maxArraySize int) (*ggml.GGML, error) {
 // NewLlamaServer will run a server for the given GPUs
 func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath string, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (LlamaServer, error) {
 	var llamaModel *llama.Model
-	var textProcessor model.TextProcessor
+	var tok tokenizer.Tokenizer
 	var err error
 	if envconfig.NewEngine() || f.KV().OllamaEngineRequired() {
 		if len(projectors) == 0 {
-			textProcessor, err = model.NewTextProcessor(modelPath)
+			tok, err = model.NewTextProcessor(modelPath)
 		} else {
 			err = errors.New("split vision models aren't supported")
 		}
@@ -154,7 +156,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 			slog.Debug("model not yet supported by Ollama engine, switching to compatibility mode", "model", modelPath, "error", err)
 		}
 	}
-	if textProcessor == nil {
+	if tok == nil {
 		llamaModel, err = llama.LoadModelFromFile(modelPath, llama.ModelParams{VocabOnly: true})
 		if err != nil {
 			return nil, err
@@ -210,7 +212,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 
 	kvct := strings.ToLower(envconfig.KvCacheType())
 
-	if textProcessor == nil {
+	if tok == nil {
 		flashAttention := ml.FlashAttentionAuto
 		if faUserSet {
 			if fa {
@@ -260,7 +262,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 	gpuLibs := ml.LibraryPaths(gpus)
 	status := NewStatusWriter(os.Stderr)
 	cmd, port, err := StartRunner(
-		textProcessor != nil,
+		tok != nil,
 		modelPath,
 		gpuLibs,
 		status,
@@ -309,8 +311,8 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 		}
 	}()
 
-	if textProcessor != nil {
-		return &ollamaServer{llmServer: s, textProcessor: textProcessor}, nil
+	if tok != nil {
+		return &ollamaServer{llmServer: s, tokenizer: tok}, nil
 	} else {
 		return &llamaServer{llmServer: s, ggml: f}, nil
 	}
@@ -1200,7 +1202,8 @@ func (s *llmServer) initModel(ctx context.Context, req LoadRequest, operation Lo
 
 	resp, err := http.DefaultClient.Do(r)
 	if err != nil {
-		return nil, fmt.Errorf("do load request: %w", err)
+		slog.Error("do load request", "error", err)
+		return nil, errors.New("model failed to load, this may be due to resource limitations or an internal error, check ollama server logs for details")
 	}
 	defer resp.Body.Close()
 
@@ -1772,7 +1775,7 @@ func (s *llamaServer) Tokenize(ctx context.Context, content string) ([]int, erro
 }
 
 func (s *ollamaServer) Tokenize(ctx context.Context, content string) ([]int, error) {
-	tokens, err := s.textProcessor.Encode(content, false)
+	tokens, err := s.tokenizer.Encode(content, false)
 	if err != nil {
 		return nil, err
 	}
@@ -1807,7 +1810,7 @@ func (s *ollamaServer) Detokenize(ctx context.Context, tokens []int) (string, er
 		toks[i] = int32(t)
 	}
 
-	content, err := s.textProcessor.Decode(toks)
+	content, err := s.tokenizer.Decode(toks)
 	if err != nil {
 		return "", err
 	}
@@ -1901,6 +1904,10 @@ func (s *llmServer) VRAMByGPU(id ml.DeviceID) uint64 {
 	return 0
 }
 
+func (s *llmServer) ContextLength() int {
+	return s.options.NumCtx
+}
+
 func (s *ollamaServer) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo {
 	devices, err := ml.GetDevicesFromRunner(ctx, s)
 	if err != nil {

manifest/layer.go

@@ -1,4 +1,4 @@
-package server
+package manifest
 
 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
+	Status    string `json:"-"`
 }
 
 const (
@@ -22,7 +22,7 @@ const (
 )
 
 func NewLayer(r io.Reader, mediatype string) (Layer, error) {
-	blobs, err := GetBlobsPath("")
+	blobs, err := BlobsPath("")
 	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 := GetBlobsPath(digest)
+	blob, err := BlobsPath(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 := GetBlobsPath(digest)
+	blob, err := BlobsPath(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 := GetBlobsPath(l.Digest)
+	blob, err := BlobsPath(l.Digest)
 	if err != nil {
 		return nil, err
 	}
@@ -126,7 +126,7 @@ func (l *Layer) Remove() error {
 		}
 	}
 
-	blob, err := GetBlobsPath(l.Digest)
+	blob, err := BlobsPath(l.Digest)
 	if err != nil {
 		return err
 	}

manifest/manifest.go

@@ -1,10 +1,9 @@
-package server
+package manifest
 
 import (
 	"crypto/sha256"
 	"encoding/hex"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -33,12 +32,38 @@ func (m *Manifest) Size() (size int64) {
 	return
 }
 
+func (m *Manifest) Digest() string {
+	return m.digest
+}
+
+func (m *Manifest) FileInfo() os.FileInfo {
+	return m.fi
+}
+
+// ReadConfigJSON reads and unmarshals a config layer as JSON.
+func (m *Manifest) ReadConfigJSON(configPath string, v any) error {
+	for _, layer := range m.Layers {
+		if layer.MediaType == "application/vnd.ollama.image.json" && layer.Name == configPath {
+			blobPath, err := BlobsPath(layer.Digest)
+			if err != nil {
+				return err
+			}
+			data, err := os.ReadFile(blobPath)
+			if err != nil {
+				return err
+			}
+			return json.Unmarshal(data, v)
+		}
+	}
+	return fmt.Errorf("config %q not found in manifest", configPath)
+}
+
 func (m *Manifest) Remove() error {
 	if err := os.Remove(m.filepath); err != nil {
 		return err
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return err
 	}
@@ -70,11 +95,11 @@ func (m *Manifest) RemoveLayers() error {
 		if _, used := inUse[layer.Digest]; used {
 			continue
 		}
-		blob, err := GetBlobsPath(layer.Digest)
+		blob, err := BlobsPath(layer.Digest)
 		if err != nil {
 			return err
 		}
-		if err := os.Remove(blob); errors.Is(err, os.ErrNotExist) {
+		if err := os.Remove(blob); os.IsNotExist(err) {
 			slog.Debug("layer does not exist", "digest", layer.Digest)
 		} else if err != nil {
 			return err
@@ -89,7 +114,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 		return nil, model.Unqualified(n)
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return nil, err
 	}
@@ -121,7 +146,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 }
 
 func WriteManifest(name model.Name, config Layer, layers []Layer) error {
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return err
 	}
@@ -148,7 +173,7 @@ func WriteManifest(name model.Name, config Layer, layers []Layer) error {
 }
 
 func Manifests(continueOnError bool) (map[model.Name]*Manifest, error) {
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return nil, err
 	}

manifest/manifest_test.go

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

manifest/paths.go

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

middleware/anthropic.go

@@ -131,12 +131,15 @@ func AnthropicMessagesMiddleware() gin.HandlerFunc {
 
 		messageID := anthropic.GenerateMessageID()
 
+		// Estimate input tokens for streaming (actual count not available until generation completes)
+		estimatedTokens := anthropic.EstimateInputTokens(req)
+
 		w := &AnthropicWriter{
 			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
 			stream:     req.Stream,
 			id:         messageID,
 			model:      req.Model,
-			converter:  anthropic.NewStreamConverter(messageID, req.Model),
+			converter:  anthropic.NewStreamConverter(messageID, req.Model, estimatedTokens),
 		}
 
 		if req.Stream {

middleware/openai.go

@@ -609,3 +609,49 @@ func ImageGenerationsMiddleware() gin.HandlerFunc {
 		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

@@ -1112,3 +1112,129 @@ func TestImageWriterResponse(t *testing.T) {
 		t.Errorf("expected image data 'dGVzdC1pbWFnZS1kYXRh', got %s", imageResp.Data[0].B64JSON)
 	}
 }
+
+func TestImageEditsMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.GenerateRequest
+		err  openai.ErrorResponse
+	}
+
+	var capturedRequest *api.GenerateRequest
+
+	// Base64-encoded test image (1x1 pixel PNG)
+	testImage := "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII="
+	decodedImage, _ := base64.StdEncoding.DecodeString("iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=")
+
+	testCases := []testCase{
+		{
+			name: "image edit basic",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue",
+				"image": "` + testImage + `"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "make it blue",
+				Images: []api.ImageData{decodedImage},
+			},
+		},
+		{
+			name: "image edit with size",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue",
+				"image": "` + testImage + `",
+				"size": "512x768"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "make it blue",
+				Images: []api.ImageData{decodedImage},
+				Width:  512,
+				Height: 768,
+			},
+		},
+		{
+			name: "image edit missing prompt",
+			body: `{
+				"model": "test-model",
+				"image": "` + testImage + `"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "prompt is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+		{
+			name: "image edit missing model",
+			body: `{
+				"prompt": "make it blue",
+				"image": "` + testImage + `"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "model is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+		{
+			name: "image edit missing image",
+			body: `{
+				"model": "test-model",
+				"prompt": "make it blue"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "image is required",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(ImageEditsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/generate", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if tc.err.Error.Message != "" {
+				var errResp openai.ErrorResponse
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+				if diff := cmp.Diff(tc.err, errResp); diff != "" {
+					t.Fatalf("errors did not match:\n%s", diff)
+				}
+				return
+			}
+
+			if resp.Code != http.StatusOK {
+				t.Fatalf("expected status 200, got %d: %s", resp.Code, resp.Body.String())
+			}
+
+			if diff := cmp.Diff(&tc.req, capturedRequest); diff != "" {
+				t.Fatalf("requests did not match:\n%s", diff)
+			}
+		})
+	}
+}

ml/backend.go

@@ -162,6 +162,7 @@ type Tensor interface {
 	AvgPool2D(ctx Context, k, s int, p float32) Tensor
 	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 	Conv3D(ctx Context, weight Tensor, c, s0, s1, s2, p0, p1, p2, d0, d1, d2 int) Tensor
+	SSMConv(ctx Context, kernel Tensor) Tensor
 
 	IM2Col(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 
@@ -169,10 +170,12 @@ type Tensor interface {
 	Cos(ctx Context) Tensor
 	Tanh(ctx Context) Tensor
 	GELU(ctx Context, up ...Tensor) Tensor
+	GELU_ERF(ctx Context) Tensor
 	QuickGELU(ctx Context, up ...Tensor) Tensor
 	SILU(ctx Context, up ...Tensor) Tensor
 	RELU(ctx Context, up ...Tensor) Tensor
 	Sigmoid(ctx Context) Tensor
+	SigmoidOut(ctx Context) Tensor
 
 	// AlphaLimitSILU is a variant of SILU that clamps the input to the range [-limit, limit]
 	SILUAlphaLimit(ctx Context, up Tensor, alpha, limit float32) Tensor
@@ -205,6 +208,32 @@ type Tensor interface {
 	Stddev(ctx Context) Tensor
 	Sqr(ctx Context) Tensor
 	Sqrt(ctx Context) Tensor
+	Exp(ctx Context) Tensor
+	Neg(ctx Context) Tensor
+
+	// Clamp clamps values to [min, max] range
+	Clamp(ctx Context, min, max float32) Tensor
+
+	// Softplus computes ln(1 + exp(x))
+	Softplus(ctx Context) Tensor
+
+	// CumSum computes cumulative sum along dimension 0
+	CumSum(ctx Context) Tensor
+
+	// Diag creates a diagonal matrix from a 1D tensor
+	Diag(ctx Context) Tensor
+
+	// Tri converts a matrix to triangular form (0=upper+diag, 1=upper, 2=lower+diag, 3=lower)
+	Tri(ctx Context, triType int) Tensor
+
+	// Fill fills a tensor with a constant value (in-place)
+	Fill(ctx Context, value float32) Tensor
+
+	// Repeat4D repeats tensor to match target shape
+	Repeat4D(ctx Context, dim0, dim1, dim2, dim3 int) Tensor
+
+	// SolveTri solves a triangular system Ax = B
+	SolveTri(ctx Context, b Tensor, lower, left, unitDiag bool) Tensor
 
 	Interpolate(ctx Context, dims [4]int, samplingMode SamplingMode) Tensor
 }

ml/backend/ggml/ggml.go

@@ -378,7 +378,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 		}
 	}
 
-	maxGraphNodes := max(1024, len(meta.Tensors().Items())*8)
+	maxGraphNodes := max(1024, len(meta.Tensors().Items())*32)
 
 	sched := C.ggml_backend_sched_new_ext(
 		(*C.ggml_backend_t)(unsafe.Pointer(&schedBackends[0])),
@@ -1468,6 +1468,13 @@ func (t *Tensor) Sigmoid(ctx ml.Context) ml.Tensor {
 	}
 }
 
+func (t *Tensor) SigmoidOut(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_sigmoid(ctx.(*Context).ctx, t.t),
+	}
+}
+
 func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	switch len(shape) {
 	case 1:
@@ -1581,6 +1588,13 @@ func (t *Tensor) GELU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
 	}
 }
 
+func (t *Tensor) GELU_ERF(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_gelu_erf_inplace(ctx.(*Context).ctx, t.t),
+	}
+}
+
 func (t *Tensor) QuickGELU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
 	var tt *C.struct_ggml_tensor
 	if len(t2) > 0 {
@@ -1641,6 +1655,13 @@ func (t *Tensor) Conv3D(ctx ml.Context, t2 ml.Tensor, c, s0, s1, s2, p0, p1, p2,
 	return tt
 }
 
+func (t *Tensor) SSMConv(ctx ml.Context, kernel ml.Tensor) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_ssm_conv(ctx.(*Context).ctx, t.t, kernel.(*Tensor).t),
+	}
+}
+
 func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
 	return &Tensor{
 		b: t.b,
@@ -1765,6 +1786,76 @@ func (t *Tensor) Sqrt(ctx ml.Context) ml.Tensor {
 	}
 }
 
+func (t *Tensor) Exp(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_exp(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) Neg(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_neg(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) Clamp(ctx ml.Context, min, max float32) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_clamp(ctx.(*Context).ctx, t.t, C.float(min), C.float(max)),
+	}
+}
+
+func (t *Tensor) Softplus(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_softplus(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) CumSum(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_cumsum(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) Diag(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_diag(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) Tri(ctx ml.Context, triType int) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_tri(ctx.(*Context).ctx, t.t, C.enum_ggml_tri_type(triType)),
+	}
+}
+
+func (t *Tensor) Fill(ctx ml.Context, value float32) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_fill_inplace(ctx.(*Context).ctx, t.t, C.float(value)),
+	}
+}
+
+func (t *Tensor) Repeat4D(ctx ml.Context, dim0, dim1, dim2, dim3 int) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_repeat_4d(ctx.(*Context).ctx, t.t, C.int64_t(dim0), C.int64_t(dim1), C.int64_t(dim2), C.int64_t(dim3)),
+	}
+}
+
+func (t *Tensor) SolveTri(ctx ml.Context, b ml.Tensor, lower, left, unitDiag bool) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_solve_tri(ctx.(*Context).ctx, t.t, b.(*Tensor).t, C._Bool(lower), C._Bool(left), C._Bool(unitDiag)),
+	}
+}
+
 func (t *Tensor) Interpolate(ctx ml.Context, dims [4]int, samplingMode ml.SamplingMode) ml.Tensor {
 	var mode C.uint32_t
 	switch samplingMode {

ml/backend/ggml/ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -66,7 +66,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  32, 128, 128, 128, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  32, 128, 128, 128, 2, true);
 
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
@@ -80,7 +81,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
 
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32,  96,  64, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
@@ -89,7 +91,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
 }
 
 static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_volta(const int DKQ, const int DV, const int ncols) {
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256,  64, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256,  64, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128,  64, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128,  64, 1, false);
@@ -397,7 +400,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     constexpr int  ncols           = ncols1 * ncols2;
     constexpr int  cols_per_warp   = T_B_KQ::I;
     constexpr int  cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
+    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2(DKQ, DV, ncols);
     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2(DKQ, DV, ncols);
@@ -467,7 +470,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 }
             }
         } else {
-            static_assert(cols_per_warp != 8, "cols_per_warp == 8 not implemented");
 #pragma unroll
             for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) {
                 load_ldmatrix(Q_B[0], tile_Q + (threadIdx.y / np)*(T_B_KQ::I*stride_tile_Q) + k_KQ_0, stride_tile_Q);
@@ -479,8 +481,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                     T_A_KQ K_A;
                     load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
 
-                    // Wide version of KQ_C is column-major => swap A and B.
-                    mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
+                    if constexpr (cols_per_warp == 8) {
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
+                    } else {
+                        // Wide version of KQ_C is column-major
+#if defined(AMD_WMMA_AVAILABLE)
+                        // RDNA matrix C is column-major.
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
+#else
+                        // swap A and B for CUDA.
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
+#endif // defined(AMD_WMMA_AVAILABLE)
+                    }
                 }
             }
         }
@@ -841,7 +853,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
 
     constexpr int  cols_per_warp   = T_B_KQ::I;
     constexpr int  cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
+    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa     (DKQ, DV, ncols);
     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2     (DKQ, DV, ncols);
     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2     (DKQ, DV, ncols);
@@ -1353,6 +1365,13 @@ static __global__ void flash_attn_ext_f16(
         NO_DEVICE_CODE;
         return;
     }
+#ifdef VOLTA_MMA_AVAILABLE
+    if (ncols1*ncols2 < 32) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif // VOLTA_MMA_AVAILABLE
+
 #if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
     if (ncols1*ncols2 > 32) {
         NO_DEVICE_CODE;
@@ -1585,3 +1604,8 @@ DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256,  64)
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);
+
+// For GLM 4.7 Flash
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  4,  4);
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  8,  4);
+extern DECL_FATTN_MMA_F16_CASE(576, 512, 16,  4);

ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cuh

@@ -68,6 +68,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
 
     return 0;
@@ -122,6 +124,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  32,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  32,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  32,  64)
 
     return 0;
@@ -183,6 +187,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128,  64)
 
@@ -245,6 +251,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128,  64)
 
@@ -1187,6 +1195,14 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
             launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
             return;
         }
+        if (use_gqa_opt && gqa_ratio % 8 == 0) {
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
+            return;
+        }
+        if (use_gqa_opt && gqa_ratio % 4 == 0) {
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
+            return;
+        }
     }
 
     if constexpr (DV <= 256) {

ml/backend/ggml/ggml/src/ggml-cuda/fattn.cu

@@ -111,7 +111,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
             ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
             break;
         case 576: {
-            // For Deepseek, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
+            // For Deepseek/GLM4, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
             GGML_ASSERT(V->ne[0] == 512);
             float max_bias = 0.0f;
             memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
@@ -121,8 +121,12 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
 
             GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
             const int gqa_ratio = Q->ne[2] / K->ne[2];
-            GGML_ASSERT(gqa_ratio % 16 == 0);
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+            GGML_ASSERT(gqa_ratio % 4 == 0);
+            if (gqa_ratio % 16 == 0) {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+            } else {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512,  4>(ctx, dst);
+            }
         } break;
         default:
             GGML_ABORT("fatal error");
@@ -251,7 +255,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
             if (V->ne[0] != 512) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!gqa_opt_applies || gqa_ratio % 16 != 0) {
+            if (!gqa_opt_applies || gqa_ratio % 4 != 0) {
                 return BEST_FATTN_KERNEL_NONE;
             }
             break;

ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4);

ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 2, 4);

ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 4, 4);

ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 8, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 8, 4);

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -1370,6 +1370,26 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm(ggml_met
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_SOLVE_TRI);
+
+    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(ggml_is_contiguous(op->src[1]));
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_solve_tri_f32");
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_group_norm(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_GROUP_NORM);
 

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.h

@@ -133,6 +133,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_norm              (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rope              (ggml_metal_library_t lib, const struct ggml_tensor * op);

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.m

@@ -1023,6 +1023,17 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             return has_simdgroup_reduction && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_L2_NORM:
             return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(op->src[0]));
+        case GGML_OP_SOLVE_TRI:
+            return ggml_is_contiguous(op->src[0]) &&
+                ggml_is_contiguous(op->src[1]) &&
+                op->src[0]->type == GGML_TYPE_F32 &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->type == GGML_TYPE_F32;
+        case GGML_OP_COUNT_EQUAL:
+            return has_simdgroup_reduction &&
+                op->src[0]->type == GGML_TYPE_I32 &&
+                op->src[1]->type == GGML_TYPE_I32 &&
+                op->type == GGML_TYPE_I64;
         case GGML_OP_ARGMAX:
             return has_simdgroup_reduction;
         case GGML_OP_NORM:
@@ -1071,12 +1082,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 op->src[0]->ne[0] != 112 &&
                 op->src[0]->ne[0] != 128 &&
                 op->src[0]->ne[0] != 192 &&
-                op->src[0]->ne[0] != 256) {
-                return false;
-            }
-            if (op->src[0]->ne[0] == 576) {
-                // DeepSeek sizes
-                // TODO: disabled for now, until optmized
+                op->src[0]->ne[0] != 256 &&
+                op->src[0]->ne[0] != 576) {
                 return false;
             }
             if (op->src[1]->type != op->src[2]->type) {

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal

@@ -2385,6 +2385,27 @@ typedef struct {
     float    eps;
 } ggml_metal_kargs_l2_norm;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_solve_tri;
+
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;
@@ -5813,6 +5834,66 @@ kernel void kernel_l2_norm_f32(
     }
 }
 
+kernel void kernel_solve_tri_f32(
+        constant ggml_metal_kargs_solve_tri & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint   tgpig[[threadgroup_position_in_grid]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    const uint64_t ncols = (uint64_t) args.ne10;
+    const uint64_t n_batches = (uint64_t) args.ne02 * (uint64_t) args.ne03;
+    const uint64_t nr = n_batches * ncols;
+
+    const uint64_t gid = (uint64_t) tgpig * (uint64_t) ntg + (uint64_t) tpitg;
+    if (gid >= nr) {
+        return;
+    }
+
+    const uint64_t i03 = gid / ((uint64_t) args.ne02 * ncols);
+    const uint64_t rem = gid - i03 * (uint64_t) args.ne02 * ncols;
+    const uint64_t i02 = rem / ncols;
+    const uint64_t i01 = rem - i02 * ncols;
+
+    const uint64_t sa0 = args.nb00 / sizeof(float);
+    const uint64_t sa1 = args.nb01 / sizeof(float);
+    const uint64_t sa2 = args.nb02 / sizeof(float);
+    const uint64_t sa3 = args.nb03 / sizeof(float);
+
+    const uint64_t sb0 = args.nb10 / sizeof(float);
+    const uint64_t sb1 = args.nb11 / sizeof(float);
+    const uint64_t sb2 = args.nb12 / sizeof(float);
+    const uint64_t sb3 = args.nb13 / sizeof(float);
+
+    const uint64_t sx0 = args.nb0 / sizeof(float);
+    const uint64_t sx1 = args.nb1 / sizeof(float);
+    const uint64_t sx2 = args.nb2 / sizeof(float);
+    const uint64_t sx3 = args.nb3 / sizeof(float);
+
+    device const float * A = (device const float *) src0;
+    device const float * B = (device const float *) src1;
+    device       float * X = (device       float *) dst;
+
+    const uint64_t A_base = i02 * sa2 + i03 * sa3;
+    const uint64_t B_base = i02 * sb2 + i03 * sb3;
+    const uint64_t X_base = i02 * sx2 + i03 * sx3;
+
+    const uint64_t n = (uint64_t) args.ne11;
+
+    for (uint64_t i00 = 0; i00 < n; ++i00) {
+        float sum = 0.0f;
+        for (uint64_t t = 0; t < i00; ++t) {
+            sum += A[A_base + i00 * sa1 + t * sa0] *
+                X[X_base + t * sx1 + i01 * sx0];
+        }
+
+        const float diag = A[A_base + i00 * sa1 + i00 * sa0];
+        X[X_base + i00 * sx1 + i01 * sx0] =
+            (B[B_base + i00 * sb1 + i01 * sb0] - sum) / diag;
+    }
+}
+
 kernel void kernel_group_norm_f32(
         constant ggml_metal_kargs_group_norm & args,
         device const float * src0,
@@ -8967,6 +9048,7 @@ kernel void kernel_flash_attn_ext(
       //case 1: kernel_flash_attn_ext_impl<FWD_TMPL, 1>(FWD_ARGS); break;
       //case 2: kernel_flash_attn_ext_impl<FWD_TMPL, 2>(FWD_ARGS); break;
         case 4: kernel_flash_attn_ext_impl<FWD_TMPL, 4>(FWD_ARGS); break;
+        case 8: kernel_flash_attn_ext_impl<FWD_TMPL, 8>(FWD_ARGS); break;
     }
 #undef FWD_TMPL
 #undef FWD_ARGS

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-impl.h

@@ -500,6 +500,27 @@ typedef struct {
     float    eps;
 } ggml_metal_kargs_l2_norm;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_solve_tri;
+
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;