server: add tests and fix isHuggingFaceURL edge case

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

CMakeLists.txt

@@ -190,7 +190,7 @@ if(MLX_ENGINE)
     install(TARGETS mlx mlxc
         RUNTIME_DEPENDENCIES
             DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
-            PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc nvrtc-builtins cudnn nccl openblas gfortran
+            PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc cudnn nccl
             PRE_EXCLUDE_REGEXES ".*"
         RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
         LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX

README.md

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

app/cmd/app/app_darwin.m

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

cmd/interactive.go

@@ -116,7 +116,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		Prompt:         ">>> ",
 		AltPrompt:      "... ",
 		Placeholder:    "Send a message (/? for help)",
-		AltPlaceholder: "Press Enter to send",
+		AltPlaceholder: `Use """ to end multi-line input`,
 	})
 	if err != nil {
 		return err

docs/api/anthropic-compatibility.mdx

@@ -21,7 +21,6 @@ ollama pull glm-4.7:cloud
 To use Ollama with tools that expect the Anthropic API (like Claude Code), set these environment variables:
 
 ```shell
-export ANTHROPIC_AUTH_TOKEN=ollama  # required but ignored
 export ANTHROPIC_BASE_URL=http://localhost:11434
 export ANTHROPIC_API_KEY=ollama  # required but ignored
 ```
@@ -248,13 +247,12 @@ curl -X POST http://localhost:11434/v1/messages \
 [Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
 
 ```shell
-ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
 ```
 
 Or set the environment variables in your shell profile:
 
 ```shell
-export ANTHROPIC_AUTH_TOKEN=ollama
 export ANTHROPIC_BASE_URL=http://localhost:11434
 export ANTHROPIC_API_KEY=ollama
 ```

docs/capabilities/web-search.mdx

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

docs/docs.json

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

docs/faq.mdx

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

docs/integrations/claude-code.mdx

@@ -25,7 +25,6 @@ Claude Code connects to Ollama using the Anthropic-compatible API.
 1. Set the environment variables:
 
 ```shell
-export ANTHROPIC_AUTH_TOKEN=ollama
 export ANTHROPIC_BASE_URL=http://localhost:11434
 export ANTHROPIC_API_KEY=ollama
 ```
@@ -39,7 +38,7 @@ claude --model qwen3-coder
 Or run with environment variables inline:
 
 ```shell
-ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
 ```
 
 ## Connecting to ollama.com

docs/integrations/marimo.mdx

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

docs/integrations/onyx.mdx

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

docs/linux.mdx

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

integration/tools_test.go

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

middleware/openai.go

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

openai/openai.go

@@ -630,10 +630,6 @@ func nameFromToolCallID(messages []Message, toolCallID string) string {
 
 // decodeImageURL decodes a base64 data URI into raw image bytes.
 func decodeImageURL(url string) (api.ImageData, error) {
-	if strings.HasPrefix(url, "http://") || strings.HasPrefix(url, "https://") {
-		return nil, errors.New("image URLs are not currently supported, please use base64 encoded data instead")
-	}
-
 	types := []string{"jpeg", "jpg", "png", "webp"}
 
 	// Support blank mime type to match /api/chat's behavior of taking just unadorned base64

openai/responses.go

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

openai/responses_test.go

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

readline/readline.go

@@ -5,7 +5,6 @@ import (
 	"fmt"
 	"io"
 	"os"
-	"strings"
 )
 
 type Prompt struct {
@@ -37,11 +36,10 @@ type Terminal struct {
 }
 
 type Instance struct {
-	Prompt      *Prompt
+	Prompt   *Prompt
-	Terminal    *Terminal
+	Terminal *Terminal
-	History     *History
+	History  *History
-	Pasting     bool
+	Pasting  bool
-	pastedLines []string
 }
 
 func New(prompt Prompt) (*Instance, error) {
@@ -176,8 +174,6 @@ func (i *Instance) Readline() (string, error) {
 		case CharEsc:
 			esc = true
 		case CharInterrupt:
-			i.pastedLines = nil
-			i.Prompt.UseAlt = false
 			return "", ErrInterrupt
 		case CharPrev:
 			i.historyPrev(buf, &currentLineBuf)
@@ -192,23 +188,7 @@ func (i *Instance) Readline() (string, error) {
 		case CharForward:
 			buf.MoveRight()
 		case CharBackspace, CharCtrlH:
-			if buf.IsEmpty() && len(i.pastedLines) > 0 {
+			buf.Remove()
-				lastIdx := len(i.pastedLines) - 1
-				prevLine := i.pastedLines[lastIdx]
-				i.pastedLines = i.pastedLines[:lastIdx]
-				fmt.Print(CursorBOL + ClearToEOL + CursorUp + CursorBOL + ClearToEOL)
-				if len(i.pastedLines) == 0 {
-					fmt.Print(i.Prompt.Prompt)
-					i.Prompt.UseAlt = false
-				} else {
-					fmt.Print(i.Prompt.AltPrompt)
-				}
-				for _, r := range prevLine {
-					buf.Add(r)
-				}
-			} else {
-				buf.Remove()
-			}
 		case CharTab:
 			// todo: convert back to real tabs
 			for range 8 {
@@ -231,28 +211,13 @@ func (i *Instance) Readline() (string, error) {
 		case CharCtrlZ:
 			fd := os.Stdin.Fd()
 			return handleCharCtrlZ(fd, i.Terminal.termios)
-		case CharCtrlJ:
+		case CharEnter, CharCtrlJ:
-			i.pastedLines = append(i.pastedLines, buf.String())
-			buf.Buf.Clear()
-			buf.Pos = 0
-			buf.DisplayPos = 0
-			buf.LineHasSpace.Clear()
-			fmt.Println()
-			fmt.Print(i.Prompt.AltPrompt)
-			i.Prompt.UseAlt = true
-			continue
-		case CharEnter:
 			output := buf.String()
-			if len(i.pastedLines) > 0 {
-				output = strings.Join(i.pastedLines, "\n") + "\n" + output
-				i.pastedLines = nil
-			}
 			if output != "" {
 				i.History.Add(output)
 			}
 			buf.MoveToEnd()
 			fmt.Println()
-			i.Prompt.UseAlt = false
 
 			return output, nil
 		default:

scripts/build_darwin.sh

@@ -179,7 +179,7 @@ _build_macapp() {
     fi
 
     rm -f dist/Ollama-darwin.zip
-    ditto -c -k --norsrc --keepParent dist/Ollama.app dist/Ollama-darwin.zip
+    ditto -c -k --keepParent dist/Ollama.app dist/Ollama-darwin.zip
     (cd dist/Ollama.app/Contents/Resources/; tar -cf - ollama ollama-mlx *.so *.dylib *.metallib 2>/dev/null) | gzip -9vc > dist/ollama-darwin.tgz
 
     # Notarize and Staple
@@ -187,7 +187,7 @@ _build_macapp() {
         $(xcrun -f notarytool) submit dist/Ollama-darwin.zip --wait --timeout 20m --apple-id "$APPLE_ID" --password "$APPLE_PASSWORD" --team-id "$APPLE_TEAM_ID"
         rm -f dist/Ollama-darwin.zip
         $(xcrun -f stapler) staple dist/Ollama.app
-        ditto -c -k --norsrc --keepParent dist/Ollama.app dist/Ollama-darwin.zip
+        ditto -c -k --keepParent dist/Ollama.app dist/Ollama-darwin.zip
 
         rm -f dist/Ollama.dmg
 

server/download.go

@@ -95,11 +95,48 @@ func (p *blobDownloadPart) UnmarshalJSON(b []byte) error {
 }
 
 const (
-	numDownloadParts          = 16
+	// numDownloadParts is the default number of concurrent download parts for standard downloads
+	numDownloadParts = 16
+	// numHFDownloadParts is the reduced number of concurrent download parts for HuggingFace
+	// downloads to avoid triggering rate limits (HTTP 429 errors). See GitHub issue #13297.
+	numHFDownloadParts        = 4
 	minDownloadPartSize int64 = 100 * format.MegaByte
 	maxDownloadPartSize int64 = 1000 * format.MegaByte
 )
 
+// isHuggingFaceURL returns true if the URL is from a HuggingFace domain.
+// This includes:
+//   - huggingface.co (main domain)
+//   - *.huggingface.co (subdomains like cdn-lfs.huggingface.co)
+//   - hf.co (shortlink domain)
+//   - *.hf.co (CDN domains like cdn-lfs.hf.co, cdn-lfs3.hf.co)
+func isHuggingFaceURL(u *url.URL) bool {
+	if u == nil {
+		return false
+	}
+	host := strings.ToLower(u.Hostname())
+	return host == "huggingface.co" ||
+		strings.HasSuffix(host, ".huggingface.co") ||
+		host == "hf.co" ||
+		strings.HasSuffix(host, ".hf.co")
+}
+
+// getNumDownloadParts returns the number of concurrent download parts to use
+// for the given URL. HuggingFace URLs use reduced concurrency (default 4) to
+// avoid triggering rate limits. This can be overridden via the OLLAMA_HF_CONCURRENCY
+// environment variable. For non-HuggingFace URLs, returns the standard concurrency (16).
+func getNumDownloadParts(u *url.URL) int {
+	if isHuggingFaceURL(u) {
+		if v := os.Getenv("OLLAMA_HF_CONCURRENCY"); v != "" {
+			if n, err := strconv.Atoi(v); err == nil && n > 0 {
+				return n
+			}
+		}
+		return numHFDownloadParts
+	}
+	return numDownloadParts
+}
+
 func (p *blobDownloadPart) Name() string {
 	return strings.Join([]string{
 		p.blobDownload.Name, "partial", strconv.Itoa(p.N),
@@ -271,7 +308,11 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 	}
 
 	g, inner := errgroup.WithContext(ctx)
-	g.SetLimit(numDownloadParts)
+	concurrency := getNumDownloadParts(directURL)
+	if concurrency != numDownloadParts {
+		slog.Info(fmt.Sprintf("using reduced concurrency (%d) for HuggingFace download", concurrency))
+	}
+	g.SetLimit(concurrency)
 	for i := range b.Parts {
 		part := b.Parts[i]
 		if part.Completed.Load() == part.Size {

server/download_test.go

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

x/README.md

@@ -0,0 +1,50 @@
+# Experimental Features
+
+## MLX Backend
+
+We're working on a new experimental backend based on the [MLX project](https://github.com/ml-explore/mlx)
+
+Support is currently limited to MacOS and Linux with CUDA GPUs. We're looking to add support for Windows CUDA soon, and other GPU vendors.
+
+### Building ollama-mlx
+
+The `ollama-mlx` binary is a separate build of Ollama with MLX support enabled. This enables experimental features like image generation.
+
+#### macOS (Apple Silicon and Intel)
+
+```bash
+# Build MLX backend libraries
+cmake --preset MLX
+cmake --build --preset MLX --parallel
+cmake --install build --component MLX
+
+# Build ollama-mlx binary
+go build -tags mlx -o ollama-mlx .
+```
+
+#### Linux (CUDA)
+
+On Linux, use the preset "MLX CUDA 13" or "MLX CUDA 12" to enable CUDA with the default Ollama NVIDIA GPU architectures enabled:
+
+```bash
+# Build MLX backend libraries with CUDA support
+cmake --preset 'MLX CUDA 13'
+cmake --build --preset 'MLX CUDA 13' --parallel
+cmake --install build --component MLX
+
+# Build ollama-mlx binary
+CGO_CFLAGS="-O3 -I$(pwd)/build/_deps/mlx-c-src" \
+CGO_LDFLAGS="-L$(pwd)/build/lib/ollama -lmlxc -lmlx" \
+go build -tags mlx -o ollama-mlx .
+```
+
+#### Using build scripts
+
+The build scripts automatically create the `ollama-mlx` binary:
+
+- **macOS**: `./scripts/build_darwin.sh` produces `dist/darwin/ollama-mlx`
+- **Linux**: `./scripts/build_linux.sh` produces `ollama-mlx` in the output archives
+
+## Image Generation
+
+Image generation is built into the `ollama-mlx` binary. Run `ollama-mlx serve` to start the server with image generation support enabled.

x/cmd/run.go

@@ -25,6 +25,14 @@ import (
 	"github.com/ollama/ollama/x/tools"
 )
 
+// MultilineState tracks the state of multiline input
+type MultilineState int
+
+const (
+	MultilineNone MultilineState = iota
+	MultilineSystem
+)
+
 // Tool output capping constants
 const (
 	// localModelTokenLimit is the token limit for local models (smaller context).
@@ -648,7 +656,7 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 		Prompt:         ">>> ",
 		AltPrompt:      "... ",
 		Placeholder:    "Send a message (/? for help)",
-		AltPlaceholder: "Press Enter to send",
+		AltPlaceholder: `Use """ to end multi-line input`,
 	})
 	if err != nil {
 		return err
@@ -699,6 +707,7 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 	var sb strings.Builder
 	var format string
 	var system string
+	var multiline MultilineState = MultilineNone
 
 	for {
 		line, err := scanner.Readline()
@@ -712,12 +721,37 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 			}
 			scanner.Prompt.UseAlt = false
 			sb.Reset()
+			multiline = MultilineNone
 			continue
 		case err != nil:
 			return err
 		}
 
 		switch {
+		case multiline != MultilineNone:
+			// check if there's a multiline terminating string
+			before, ok := strings.CutSuffix(line, `"""`)
+			sb.WriteString(before)
+			if !ok {
+				fmt.Fprintln(&sb)
+				continue
+			}
+
+			switch multiline {
+			case MultilineSystem:
+				system = sb.String()
+				newMessage := api.Message{Role: "system", Content: system}
+				if len(messages) > 0 && messages[len(messages)-1].Role == "system" {
+					messages[len(messages)-1] = newMessage
+				} else {
+					messages = append(messages, newMessage)
+				}
+				fmt.Println("Set system message.")
+				sb.Reset()
+			}
+
+			multiline = MultilineNone
+			scanner.Prompt.UseAlt = false
 		case strings.HasPrefix(line, "/exit"), strings.HasPrefix(line, "/bye"):
 			return nil
 		case strings.HasPrefix(line, "/clear"):
@@ -826,18 +860,41 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 					options[args[2]] = fp[args[2]]
 				case "system":
 					if len(args) < 3 {
-						fmt.Println("Usage: /set system <message>")
+						fmt.Println("Usage: /set system <message> or /set system \"\"\"<multi-line message>\"\"\"")
 						continue
 					}
 
-					system = strings.Join(args[2:], " ")
+					multiline = MultilineSystem
-					newMessage := api.Message{Role: "system", Content: system}
+
+					line := strings.Join(args[2:], " ")
+					line, ok := strings.CutPrefix(line, `"""`)
+					if !ok {
+						multiline = MultilineNone
+					} else {
+						// only cut suffix if the line is multiline
+						line, ok = strings.CutSuffix(line, `"""`)
+						if ok {
+							multiline = MultilineNone
+						}
+					}
+
+					sb.WriteString(line)
+					if multiline != MultilineNone {
+						scanner.Prompt.UseAlt = true
+						continue
+					}
+
+					system = sb.String()
+					newMessage := api.Message{Role: "system", Content: sb.String()}
+					// Check if the slice is not empty and the last message is from 'system'
 					if len(messages) > 0 && messages[len(messages)-1].Role == "system" {
+						// Replace the last message
 						messages[len(messages)-1] = newMessage
 					} else {
 						messages = append(messages, newMessage)
 					}
 					fmt.Println("Set system message.")
+					sb.Reset()
 					continue
 				default:
 					fmt.Printf("Unknown command '/set %s'. Type /? for help\n", args[1])
@@ -1024,7 +1081,7 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 			sb.WriteString(line)
 		}
 
-		if sb.Len() > 0 {
+		if sb.Len() > 0 && multiline == MultilineNone {
 			newMessage := api.Message{Role: "user", Content: sb.String()}
 			messages = append(messages, newMessage)