cmd: set context limits for cloud models in opencode (#14107)

Allow installing Ollama on MacOS directly from the command line. This is in line with other CLI tools and results in a more streamlined experience when the user is looking to use the CLI specifically.

.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

anthropic/anthropic.go

@@ -518,24 +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),
 	}
 }
 
@@ -551,7 +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",
@@ -779,3 +785,123 @@ 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)
+		}
+	}
+
+	if source, ok := blockMap["source"].(map[string]any); ok {
+		if data, ok := source["data"].(string); ok {
+			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{
@@ -678,7 +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",
@@ -731,7 +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)
@@ -778,7 +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()
@@ -842,10 +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
@@ -899,11 +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",
@@ -937,7 +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",
@@ -969,3 +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)
+}

cmd/cmd.go

@@ -367,14 +367,25 @@ 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
 	}
 
@@ -1752,7 +1763,7 @@ 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

cmd/cmd_test.go

@@ -3,6 +3,7 @@ package cmd
 import (
 	"bytes"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
@@ -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

@@ -1,18 +1,23 @@
 package config
 
 import (
+	"context"
 	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"runtime"
 
+	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 )
 
-// Claude implements Runner for Claude Code integration
+// 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 {
@@ -60,3 +65,104 @@ func (c *Claude) Run(model string, args []string) error {
 	)
 	return cmd.Run()
 }
+
+// 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 := selectPrompt("Select model:", items)
+		fmt.Fprintf(os.Stderr, "\033[3A\033[J")
+		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/config.go

@@ -13,7 +13,8 @@ import (
 )
 
 type integration struct {
-	Models []string `json:"models"`
+	Models  []string          `json:"models"`
+	Aliases map[string]string `json:"aliases,omitempty"`
 }
 
 type config struct {
@@ -133,8 +134,16 @@ func saveIntegration(appName string, models []string) error {
 		return err
 	}
 
-	cfg.Integrations[strings.ToLower(appName)] = &integration{
-		Models: models,
+	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)
@@ -154,6 +163,29 @@ func loadIntegration(appName string) (*integration, error) {
 	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 {

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

@@ -46,6 +46,53 @@ func TestIntegrationConfig(t *testing.T) {
 		}
 	})
 
+	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"})
 

cmd/config/integrations.go

@@ -39,6 +39,15 @@ type Editor interface {
 	Models() []string
 }
 
+// AliasConfigurer can configure model aliases (e.g., for subagent routing).
+// Integrations like Claude and Codex use this to route model requests to local models.
+type AliasConfigurer interface {
+	// ConfigureAliases prompts the user to configure aliases and returns the updated map.
+	ConfigureAliases(ctx context.Context, primaryModel string, existing map[string]string, force bool) (map[string]string, bool, error)
+	// SetAliases syncs the configured aliases to the server
+	SetAliases(ctx context.Context, aliases map[string]string) error
+}
+
 // integrations is the registry of available integrations.
 var integrations = map[string]Runner{
 	"claude":   &Claude{},
@@ -48,7 +57,6 @@ var integrations = map[string]Runner{
 	"droid":    &Droid{},
 	"opencode": &OpenCode{},
 	"openclaw": &Openclaw{},
-	"pi":       &Pi{},
 }
 
 // recommendedModels are shown when the user has no models or as suggestions.
@@ -64,7 +72,6 @@ var recommendedModels = []selectItem{
 var integrationAliases = map[string]bool{
 	"clawdbot": true,
 	"moltbot":  true,
-	"pi":       true,
 }
 
 func selectIntegration() (string, error) {
@@ -131,7 +138,11 @@ func selectModels(ctx context.Context, name, current string) ([]string, error) {
 			return nil, err
 		}
 	} else {
-		model, err := selectPrompt(fmt.Sprintf("Select model for %s:", r), items)
+		prompt := fmt.Sprintf("Select model for %s:", r)
+		if _, ok := r.(AliasConfigurer); ok {
+			prompt = fmt.Sprintf("Select Primary model for %s:", r)
+		}
+		model, err := selectPrompt(prompt, items)
 		if err != nil {
 			return nil, err
 		}
@@ -159,73 +170,123 @@ func selectModels(ctx context.Context, name, current string) ([]string, error) {
 		}
 	}
 
+	if err := ensureAuth(ctx, client, cloudModels, selected); err != nil {
+		return nil, err
+	}
+
+	return selected, nil
+}
+
+func pullIfNeeded(ctx context.Context, client *api.Client, existingModels map[string]bool, model string) error {
+	if existingModels[model] {
+		return nil
+	}
+	msg := fmt.Sprintf("Download %s?", model)
+	if ok, err := confirmPrompt(msg); err != nil {
+		return err
+	} else if !ok {
+		return errCancelled
+	}
+	fmt.Fprintf(os.Stderr, "\n")
+	if err := pullModel(ctx, client, model); err != nil {
+		return fmt.Errorf("failed to pull %s: %w", model, err)
+	}
+	return nil
+}
+
+func listModels(ctx context.Context) ([]selectItem, map[string]bool, map[string]bool, *api.Client, error) {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return nil, nil, nil, nil, err
+	}
+
+	models, err := client.List(ctx)
+	if err != nil {
+		return nil, nil, nil, nil, err
+	}
+
+	var existing []modelInfo
+	for _, m := range models.Models {
+		existing = append(existing, modelInfo{
+			Name:   m.Name,
+			Remote: m.RemoteModel != "",
+		})
+	}
+
+	items, _, existingModels, cloudModels := buildModelList(existing, nil, "")
+
+	if len(items) == 0 {
+		return nil, nil, nil, nil, fmt.Errorf("no models available, run 'ollama pull <model>' first")
+	}
+
+	return items, existingModels, cloudModels, client, nil
+}
+
+func ensureAuth(ctx context.Context, client *api.Client, cloudModels map[string]bool, selected []string) error {
 	var selectedCloudModels []string
 	for _, m := range selected {
 		if cloudModels[m] {
 			selectedCloudModels = append(selectedCloudModels, m)
 		}
 	}
-	if len(selectedCloudModels) > 0 {
-		// ensure user is signed in
-		user, err := client.Whoami(ctx)
-		if err == nil && user != nil && user.Name != "" {
-			return selected, nil
-		}
+	if len(selectedCloudModels) == 0 {
+		return nil
+	}
 
-		var aErr api.AuthorizationError
-		if !errors.As(err, &aErr) || aErr.SigninURL == "" {
-			return nil, err
-		}
+	user, err := client.Whoami(ctx)
+	if err == nil && user != nil && user.Name != "" {
+		return nil
+	}
 
-		modelList := strings.Join(selectedCloudModels, ", ")
-		yes, err := confirmPrompt(fmt.Sprintf("sign in to use %s?", modelList))
-		if err != nil || !yes {
-			return nil, fmt.Errorf("%s requires sign in", modelList)
-		}
+	var aErr api.AuthorizationError
+	if !errors.As(err, &aErr) || aErr.SigninURL == "" {
+		return err
+	}
 
-		fmt.Fprintf(os.Stderr, "\nTo sign in, navigate to:\n    %s\n\n", aErr.SigninURL)
+	modelList := strings.Join(selectedCloudModels, ", ")
+	yes, err := confirmPrompt(fmt.Sprintf("sign in to use %s?", modelList))
+	if err != nil || !yes {
+		return fmt.Errorf("%s requires sign in", modelList)
+	}
 
-		// TODO(parthsareen): extract into auth package for cmd
-		// Auto-open browser (best effort, fail silently)
-		switch runtime.GOOS {
-		case "darwin":
-			_ = exec.Command("open", aErr.SigninURL).Start()
-		case "linux":
-			_ = exec.Command("xdg-open", aErr.SigninURL).Start()
-		case "windows":
-			_ = exec.Command("rundll32", "url.dll,FileProtocolHandler", aErr.SigninURL).Start()
-		}
+	fmt.Fprintf(os.Stderr, "\nTo sign in, navigate to:\n    %s\n\n", aErr.SigninURL)
 
-		spinnerFrames := []string{"|", "/", "-", "\\"}
-		frame := 0
+	switch runtime.GOOS {
+	case "darwin":
+		_ = exec.Command("open", aErr.SigninURL).Start()
+	case "linux":
+		_ = exec.Command("xdg-open", aErr.SigninURL).Start()
+	case "windows":
+		_ = exec.Command("rundll32", "url.dll,FileProtocolHandler", aErr.SigninURL).Start()
+	}
 
-		fmt.Fprintf(os.Stderr, "\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[0])
+	spinnerFrames := []string{"|", "/", "-", "\\"}
+	frame := 0
 
-		ticker := time.NewTicker(200 * time.Millisecond)
-		defer ticker.Stop()
+	fmt.Fprintf(os.Stderr, "\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[0])
 
-		for {
-			select {
-			case <-ctx.Done():
-				fmt.Fprintf(os.Stderr, "\r\033[K")
-				return nil, ctx.Err()
-			case <-ticker.C:
-				frame++
-				fmt.Fprintf(os.Stderr, "\r\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[frame%len(spinnerFrames)])
+	ticker := time.NewTicker(200 * time.Millisecond)
+	defer ticker.Stop()
 
-				// poll every 10th frame (~2 seconds)
-				if frame%10 == 0 {
-					u, err := client.Whoami(ctx)
-					if err == nil && u != nil && u.Name != "" {
-						fmt.Fprintf(os.Stderr, "\r\033[K\033[A\r\033[K\033[1msigned in:\033[0m %s\n", u.Name)
-						return selected, nil
-					}
+	for {
+		select {
+		case <-ctx.Done():
+			fmt.Fprintf(os.Stderr, "\r\033[K")
+			return ctx.Err()
+		case <-ticker.C:
+			frame++
+			fmt.Fprintf(os.Stderr, "\r\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[frame%len(spinnerFrames)])
+
+			// poll every 10th frame (~2 seconds)
+			if frame%10 == 0 {
+				u, err := client.Whoami(ctx)
+				if err == nil && u != nil && u.Name != "" {
+					fmt.Fprintf(os.Stderr, "\r\033[K\033[A\r\033[K\033[1msigned in:\033[0m %s\n", u.Name)
+					return nil
 				}
 			}
 		}
 	}
-
-	return selected, nil
 }
 
 func runIntegration(name, modelName string, args []string) error {
@@ -233,10 +294,33 @@ func runIntegration(name, modelName string, args []string) error {
 	if !ok {
 		return fmt.Errorf("unknown integration: %s", name)
 	}
+
 	fmt.Fprintf(os.Stderr, "\nLaunching %s with %s...\n", r, modelName)
 	return r.Run(modelName, args)
 }
 
+// syncAliases syncs aliases to server and saves locally for an AliasConfigurer.
+func syncAliases(ctx context.Context, client *api.Client, ac AliasConfigurer, name, model string, existing map[string]string) error {
+	aliases := make(map[string]string)
+	for k, v := range existing {
+		aliases[k] = v
+	}
+	aliases["primary"] = model
+
+	if isCloudModel(ctx, client, model) {
+		if aliases["fast"] == "" || !isCloudModel(ctx, client, aliases["fast"]) {
+			aliases["fast"] = model
+		}
+	} else {
+		delete(aliases, "fast")
+	}
+
+	if err := ac.SetAliases(ctx, aliases); err != nil {
+		return err
+	}
+	return saveAliases(name, aliases)
+}
+
 // LaunchCmd returns the cobra command for launching integrations.
 func LaunchCmd(checkServerHeartbeat func(cmd *cobra.Command, args []string) error) *cobra.Command {
 	var modelFlag string
@@ -304,9 +388,87 @@ Examples:
 				return fmt.Errorf("unknown integration: %s", name)
 			}
 
-			if !configFlag && modelFlag == "" {
-				if config, err := loadIntegration(name); err == nil && len(config.Models) > 0 {
-					return runIntegration(name, config.Models[0], passArgs)
+			// Handle AliasConfigurer integrations (claude, codex)
+			if ac, ok := r.(AliasConfigurer); ok {
+				client, err := api.ClientFromEnvironment()
+				if err != nil {
+					return err
+				}
+
+				// Validate --model flag if provided
+				if modelFlag != "" {
+					if _, err := client.Show(cmd.Context(), &api.ShowRequest{Name: modelFlag}); err != nil {
+						return fmt.Errorf("model %q not found", modelFlag)
+					}
+				}
+
+				var model string
+				var existingAliases map[string]string
+
+				// Load saved config
+				if cfg, err := loadIntegration(name); err == nil {
+					existingAliases = cfg.Aliases
+					if len(cfg.Models) > 0 {
+						model = cfg.Models[0]
+						// AliasConfigurer integrations use single model; sanitize if multiple
+						if len(cfg.Models) > 1 {
+							_ = saveIntegration(name, []string{model})
+						}
+					}
+				}
+
+				// --model flag overrides saved model
+				if modelFlag != "" {
+					model = modelFlag
+				}
+
+				// Validate saved model still exists
+				if model != "" && modelFlag == "" {
+					if _, err := client.Show(cmd.Context(), &api.ShowRequest{Name: model}); err != nil {
+						fmt.Fprintf(os.Stderr, "%sConfigured model %q not found%s\n\n", ansiGray, model, ansiReset)
+						model = ""
+					}
+				}
+
+				// If no valid model or --config flag, show picker
+				if model == "" || configFlag {
+					aliases, _, err := ac.ConfigureAliases(cmd.Context(), model, existingAliases, configFlag)
+					if errors.Is(err, errCancelled) {
+						return nil
+					}
+					if err != nil {
+						return err
+					}
+					model = aliases["primary"]
+					existingAliases = aliases
+				}
+
+				// Sync aliases and save
+				if err := syncAliases(cmd.Context(), client, ac, name, model, existingAliases); err != nil {
+					fmt.Fprintf(os.Stderr, "%sWarning: Could not sync aliases: %v%s\n", ansiGray, err, ansiReset)
+				}
+				if err := saveIntegration(name, []string{model}); err != nil {
+					return fmt.Errorf("failed to save: %w", err)
+				}
+
+				// Launch (unless --config without confirmation)
+				if configFlag {
+					if launch, _ := confirmPrompt(fmt.Sprintf("Launch %s now?", r)); launch {
+						return runIntegration(name, model, passArgs)
+					}
+					return nil
+				}
+				return runIntegration(name, model, passArgs)
+			}
+
+			// Validate --model flag for non-AliasConfigurer integrations
+			if modelFlag != "" {
+				client, err := api.ClientFromEnvironment()
+				if err != nil {
+					return err
+				}
+				if _, err := client.Show(cmd.Context(), &api.ShowRequest{Name: modelFlag}); err != nil {
+					return fmt.Errorf("model %q not found", modelFlag)
 				}
 			}
 
@@ -320,6 +482,8 @@ Examples:
 						}
 					}
 				}
+			} else if saved, err := loadIntegration(name); err == nil && len(saved.Models) > 0 && !configFlag {
+				return runIntegration(name, saved.Models[0], passArgs)
 			} else {
 				var err error
 				models, err = selectModels(cmd.Context(), name, "")
@@ -382,8 +546,9 @@ Examples:
 }
 
 type modelInfo struct {
-	Name   string
-	Remote bool
+	Name        string
+	Remote      bool
+	ToolCapable bool
 }
 
 // buildModelList merges existing models with recommendations, sorts them, and returns
@@ -420,7 +585,7 @@ func buildModelList(existing []modelInfo, preChecked []string, current string) (
 			continue
 		}
 		items = append(items, rec)
-		if isCloudModel(rec.Name) {
+		if strings.HasSuffix(rec.Name, ":cloud") {
 			cloudModels[rec.Name] = true
 		}
 	}
@@ -480,8 +645,16 @@ func buildModelList(existing []modelInfo, preChecked []string, current string) (
 	return items, preChecked, existingModels, cloudModels
 }
 
-func isCloudModel(name string) bool {
-	return strings.HasSuffix(name, ":cloud")
+// isCloudModel checks if a model is a cloud model using the Show API.
+func isCloudModel(ctx context.Context, client *api.Client, name string) bool {
+	if client == nil {
+		return false
+	}
+	resp, err := client.Show(ctx, &api.ShowRequest{Name: name})
+	if err != nil {
+		return false
+	}
+	return resp.RemoteModel != ""
 }
 
 func pullModel(ctx context.Context, client *api.Client, model string) error {

cmd/config/integrations_test.go

@@ -1,6 +1,7 @@
 package config
 
 import (
+	"context"
 	"fmt"
 	"slices"
 	"strings"
@@ -297,24 +298,15 @@ func TestParseArgs(t *testing.T) {
 }
 
 func TestIsCloudModel(t *testing.T) {
-	tests := []struct {
-		name string
-		want bool
-	}{
-		{"glm-4.7:cloud", true},
-		{"kimi-k2.5:cloud", true},
-		{"glm-4.7-flash", false},
-		{"glm-4.7-flash:latest", false},
-		{"cloud-model", false},
-		{"model:cloudish", false},
-	}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			if got := isCloudModel(tt.name); got != tt.want {
-				t.Errorf("isCloudModel(%q) = %v, want %v", tt.name, got, tt.want)
+	// isCloudModel now only uses Show API, so nil client always returns false
+	t.Run("nil client returns false", func(t *testing.T) {
+		models := []string{"glm-4.7:cloud", "kimi-k2.5:cloud", "local-model"}
+		for _, model := range models {
+			if isCloudModel(context.Background(), nil, model) {
+				t.Errorf("isCloudModel(%q) with nil client should return false", model)
 			}
-		})
-	}
+		}
+	})
 }
 
 func names(items []selectItem) []string {
@@ -509,3 +501,41 @@ func TestBuildModelList_ReturnsExistingAndCloudMaps(t *testing.T) {
 		t.Error("llama3.2 should not be in cloudModels")
 	}
 }
+
+func TestEditorIntegration_SavedConfigSkipsSelection(t *testing.T) {
+	tmpDir := t.TempDir()
+	setTestHome(t, tmpDir)
+
+	// Save a config for opencode so it looks like a previous launch
+	if err := saveIntegration("opencode", []string{"llama3.2"}); err != nil {
+		t.Fatal(err)
+	}
+
+	// Verify loadIntegration returns the saved models
+	saved, err := loadIntegration("opencode")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(saved.Models) == 0 {
+		t.Fatal("expected saved models")
+	}
+	if saved.Models[0] != "llama3.2" {
+		t.Errorf("expected llama3.2, got %s", saved.Models[0])
+	}
+}
+
+func TestAliasConfigurerInterface(t *testing.T) {
+	t.Run("claude implements AliasConfigurer", func(t *testing.T) {
+		claude := &Claude{}
+		if _, ok := interface{}(claude).(AliasConfigurer); !ok {
+			t.Error("Claude should implement AliasConfigurer")
+		}
+	})
+
+	t.Run("codex does not implement AliasConfigurer", func(t *testing.T) {
+		codex := &Codex{}
+		if _, ok := interface{}(codex).(AliasConfigurer); ok {
+			t.Error("Codex should not implement AliasConfigurer")
+		}
+	})
+}

cmd/config/openclaw.go

@@ -17,8 +17,6 @@ type Openclaw struct{}
 
 func (c *Openclaw) String() string { return "OpenClaw" }
 
-const ansiGreen = "\033[32m"
-
 func (c *Openclaw) Run(model string, args []string) error {
 	bin := "openclaw"
 	if _, err := exec.LookPath(bin); err != nil {

cmd/config/opencode.go

@@ -1,6 +1,7 @@
 package config
 
 import (
+	"context"
 	"encoding/json"
 	"fmt"
 	"maps"
@@ -10,12 +11,52 @@ import (
 	"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-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 {
@@ -113,6 +154,8 @@ func (o *OpenCode) Edit(modelList []string) error {
 		}
 	}
 
+	client, _ := api.ClientFromEnvironment()
+
 	for _, model := range modelList {
 		if existing, ok := models[model].(map[string]any); ok {
 			// migrate existing models without _launch marker
@@ -122,12 +165,29 @@ func (o *OpenCode) Edit(modelList []string) error {
 					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
 		}
-		models[model] = map[string]any{
+		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

cmd/config/opencode_test.go

@@ -2,6 +2,7 @@ package config
 
 import (
 	"encoding/json"
+	"fmt"
 	"os"
 	"path/filepath"
 	"testing"
@@ -495,6 +496,165 @@ func TestOpenCodeEdit_SpecialCharsInModelName(t *testing.T) {
 	}
 }
 
+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},
+		{"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()

cmd/config/pi.go

@@ -1,196 +0,0 @@
-package config
-
-import (
-	"encoding/json"
-	"fmt"
-	"os"
-	"os/exec"
-	"path/filepath"
-	"slices"
-
-	"github.com/ollama/ollama/envconfig"
-)
-
-// 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
-	for _, model := range models {
-		if selectedSet[model] {
-			newModels = append(newModels, map[string]any{
-				"id":      model,
-				"_launch": true,
-			})
-		}
-	}
-
-	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
-}

cmd/config/pi_test.go

@@ -1,609 +0,0 @@
-package config
-
-import (
-	"encoding/json"
-	"os"
-	"path/filepath"
-	"testing"
-)
-
-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) {
-	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)
-		}
-	})
-}

cmd/config/selector.go

@@ -17,6 +17,7 @@ const (
 	ansiBold       = "\033[1m"
 	ansiReset      = "\033[0m"
 	ansiGray       = "\033[37m"
+	ansiGreen      = "\033[32m"
 	ansiClearDown  = "\033[J"
 )
 

cmd/config/selector_test.go

@@ -96,6 +96,14 @@ func TestSelectState(t *testing.T) {
 		}
 	})
 
+	t.Run("Enter_EmptyFilteredList_EmptyFilter_DoesNothing", func(t *testing.T) {
+		s := newSelectState([]selectItem{})
+		done, result, err := s.handleInput(eventEnter, 0)
+		if done || result != "" || err != nil {
+			t.Errorf("expected (false, '', nil), got (%v, %v, %v)", done, result, err)
+		}
+	})
+
 	t.Run("Escape_ReturnsCancelledError", func(t *testing.T) {
 		s := newSelectState(items)
 		done, result, err := s.handleInput(eventEscape, 0)
@@ -574,8 +582,19 @@ func TestRenderSelect(t *testing.T) {
 		var buf bytes.Buffer
 		renderSelect(&buf, "Select:", s)
 
+		output := buf.String()
+		if !strings.Contains(output, "no matches") {
+			t.Errorf("expected 'no matches' message, got: %s", output)
+		}
+	})
+
+	t.Run("EmptyFilteredList_EmptyFilter_ShowsNoMatches", func(t *testing.T) {
+		s := newSelectState([]selectItem{})
+		var buf bytes.Buffer
+		renderSelect(&buf, "Select:", s)
+
 		if !strings.Contains(buf.String(), "no matches") {
-			t.Error("expected 'no matches' message")
+			t.Error("expected 'no matches' message for empty list with no filter")
 		}
 	})
 

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

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)
+	}
+}

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 {

ml/backend.go

@@ -175,6 +175,7 @@ type Tensor interface {
 	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

ml/backend/ggml/ggml.go

@@ -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:

model/models/qwen3next/model.go

@@ -135,7 +135,7 @@ func (mlp *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options
 		// Apply shared expert gating
 		if mlp.SharedGateInp != nil {
 			sharedGateVal := mlp.SharedGateInp.Forward(ctx, hiddenStates2D)
-			sharedGateVal = sharedGateVal.Sigmoid(ctx)
+			sharedGateVal = sharedGateVal.SigmoidOut(ctx)
 			// Broadcast gate to match dimensions
 			sharedGateVal = sharedGateVal.Repeat(ctx, 0, sharedOut.Dim(0))
 			sharedOut = sharedOut.Mul(ctx, sharedGateVal)

runner/ollamarunner/runner.go

@@ -514,13 +514,6 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 			continue
 		}
 
-		// if past the num predict limit
-		if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
-			s.removeSequence(seqIdx, llm.DoneReasonLength)
-			nextBatch.seqs[seqIdx] = nil
-			continue
-		}
-
 		if !s.cache.enabled {
 			seq.inputs = append(seq.cache.Inputs, seq.inputs...)
 			seq.cache.Inputs = []*input.Input{}
@@ -709,7 +702,6 @@ func (s *Server) computeBatch(activeBatch batchState) {
 			continue
 		}
 
-		seq.numPredicted++
 		nextToken := &input.Input{Token: 0} // placeholder we'll fill in after Compute/Floats
 		seq.inputs = []*input.Input{nextToken}
 		nextBatchTokens[i] = nextToken
@@ -745,7 +737,9 @@ func (s *Server) computeBatch(activeBatch batchState) {
 			logutil.Trace("computeBatch: sequence replaced, discarding its results", "batchID", activeBatch.id, "seqIdx", i)
 			continue
 		}
+
 		seq.lastUpdatedAt = t
+		seq.numPredicted++
 		if seq.numPredicted == 1 {
 			seq.processingDuration = seq.lastUpdatedAt.Sub(seq.startedAt)
 			seq.startedAt = seq.lastUpdatedAt
@@ -791,6 +785,13 @@ func (s *Server) computeBatch(activeBatch batchState) {
 		}
 
 		seq.pendingResponses = append(seq.pendingResponses, piece)
+
+		// if past the num predict limit
+		if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
+			s.removeSequence(i, llm.DoneReasonLength)
+			continue
+		}
+
 		sequence := strings.Join(seq.pendingResponses, "")
 
 		if ok, stop := common.FindStop(sequence, seq.stop); ok {

scripts/install.sh

@@ -1,5 +1,5 @@
 #!/bin/sh
-# This script installs Ollama on Linux.
+# This script installs Ollama on Linux and macOS.
 # It detects the current operating system architecture and installs the appropriate version of Ollama.
 
 set -eu
@@ -27,8 +27,7 @@ require() {
     echo $MISSING
 }
 
-[ "$(uname -s)" = "Linux" ] || error 'This script is intended to run on Linux only.'
-
+OS="$(uname -s)"
 ARCH=$(uname -m)
 case "$ARCH" in
     x86_64) ARCH="amd64" ;;
@@ -36,6 +35,65 @@ case "$ARCH" in
     *) error "Unsupported architecture: $ARCH" ;;
 esac
 
+###########################################
+# macOS
+###########################################
+
+if [ "$OS" = "Darwin" ]; then
+    NEEDS=$(require curl unzip)
+    if [ -n "$NEEDS" ]; then
+        status "ERROR: The following tools are required but missing:"
+        for NEED in $NEEDS; do
+            echo "  - $NEED"
+        done
+        exit 1
+    fi
+
+    if [ -n "${OLLAMA_VERSION:-}" ]; then
+        DOWNLOAD_URL="https://github.com/ollama/ollama/releases/download/${OLLAMA_VERSION}/Ollama-darwin.zip"
+    else
+        DOWNLOAD_URL="https://github.com/ollama/ollama/releases/latest/download/Ollama-darwin.zip"
+    fi
+
+    if pgrep -x Ollama >/dev/null 2>&1; then
+        status "Stopping running Ollama instance..."
+        pkill -x Ollama 2>/dev/null || true
+        sleep 2
+    fi
+
+    if [ -d "/Applications/Ollama.app" ]; then
+        status "Removing existing Ollama installation..."
+        rm -rf "/Applications/Ollama.app"
+    fi
+
+    status "Downloading Ollama for macOS..."
+    curl --fail --show-error --location --progress-bar \
+        -o "$TEMP_DIR/Ollama-darwin.zip" "$DOWNLOAD_URL"
+
+    status "Installing Ollama to /Applications..."
+    unzip -q "$TEMP_DIR/Ollama-darwin.zip" -d "$TEMP_DIR"
+    mv "$TEMP_DIR/Ollama.app" "/Applications/"
+
+    status "Adding 'ollama' command to PATH (may require password)..."
+    mkdir -p "/usr/local/bin" 2>/dev/null || sudo mkdir -p "/usr/local/bin"
+    ln -sf "/Applications/Ollama.app/Contents/Resources/ollama" "/usr/local/bin/ollama" 2>/dev/null || \
+        sudo ln -sf "/Applications/Ollama.app/Contents/Resources/ollama" "/usr/local/bin/ollama"
+
+    if [ -z "${OLLAMA_NO_START:-}" ]; then
+        status "Starting Ollama..."
+        open -a Ollama --args hidden
+    fi
+
+    status "Install complete. You can now run 'ollama'."
+    exit 0
+fi
+
+###########################################
+# Linux
+###########################################
+
+[ "$OS" = "Linux" ] || error 'This script is intended to run on Linux and macOS only.'
+
 IS_WSL2=false
 
 KERN=$(uname -r)

server/aliases.go

@@ -0,0 +1,422 @@
+package server
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"sync"
+
+	"github.com/ollama/ollama/manifest"
+	"github.com/ollama/ollama/types/model"
+)
+
+const (
+	serverConfigFilename = "server.json"
+	serverConfigVersion  = 1
+)
+
+var errAliasCycle = errors.New("alias cycle detected")
+
+type aliasEntry struct {
+	Alias          string `json:"alias"`
+	Target         string `json:"target"`
+	PrefixMatching bool   `json:"prefix_matching,omitempty"`
+}
+
+type serverConfig struct {
+	Version int          `json:"version"`
+	Aliases []aliasEntry `json:"aliases"`
+}
+
+type store struct {
+	mu            sync.RWMutex
+	path          string
+	entries       map[string]aliasEntry // normalized alias -> entry (exact matches)
+	prefixEntries []aliasEntry          // prefix matches, sorted longest-first
+}
+
+func createStore(path string) (*store, error) {
+	store := &store{
+		path:    path,
+		entries: make(map[string]aliasEntry),
+	}
+	if err := store.load(); err != nil {
+		return nil, err
+	}
+	return store, nil
+}
+
+func (s *store) load() error {
+	data, err := os.ReadFile(s.path)
+	if err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			return nil
+		}
+		return err
+	}
+
+	var cfg serverConfig
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return err
+	}
+
+	if cfg.Version != 0 && cfg.Version != serverConfigVersion {
+		return fmt.Errorf("unsupported router config version %d", cfg.Version)
+	}
+
+	for _, entry := range cfg.Aliases {
+		targetName := model.ParseName(entry.Target)
+		if !targetName.IsValid() {
+			slog.Warn("invalid alias target in router config", "target", entry.Target)
+			continue
+		}
+		canonicalTarget := displayAliasName(targetName)
+
+		if entry.PrefixMatching {
+			// Prefix aliases don't need to be valid model names
+			alias := strings.TrimSpace(entry.Alias)
+			if alias == "" {
+				slog.Warn("empty prefix alias in router config")
+				continue
+			}
+			s.prefixEntries = append(s.prefixEntries, aliasEntry{
+				Alias:          alias,
+				Target:         canonicalTarget,
+				PrefixMatching: true,
+			})
+		} else {
+			aliasName := model.ParseName(entry.Alias)
+			if !aliasName.IsValid() {
+				slog.Warn("invalid alias name in router config", "alias", entry.Alias)
+				continue
+			}
+			canonicalAlias := displayAliasName(aliasName)
+			s.entries[normalizeAliasKey(aliasName)] = aliasEntry{
+				Alias:  canonicalAlias,
+				Target: canonicalTarget,
+			}
+		}
+	}
+
+	// Sort prefix entries by alias length descending (longest prefix wins)
+	s.sortPrefixEntriesLocked()
+
+	return nil
+}
+
+func (s *store) saveLocked() error {
+	dir := filepath.Dir(s.path)
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return err
+	}
+
+	// Combine exact and prefix entries
+	entries := make([]aliasEntry, 0, len(s.entries)+len(s.prefixEntries))
+	for _, entry := range s.entries {
+		entries = append(entries, entry)
+	}
+	entries = append(entries, s.prefixEntries...)
+
+	sort.Slice(entries, func(i, j int) bool {
+		return strings.Compare(entries[i].Alias, entries[j].Alias) < 0
+	})
+
+	cfg := serverConfig{
+		Version: serverConfigVersion,
+		Aliases: entries,
+	}
+
+	f, err := os.CreateTemp(dir, "router-*.json")
+	if err != nil {
+		return err
+	}
+
+	enc := json.NewEncoder(f)
+	enc.SetIndent("", "  ")
+	if err := enc.Encode(cfg); err != nil {
+		_ = f.Close()
+		_ = os.Remove(f.Name())
+		return err
+	}
+
+	if err := f.Close(); err != nil {
+		_ = os.Remove(f.Name())
+		return err
+	}
+
+	if err := os.Chmod(f.Name(), 0o644); err != nil {
+		_ = os.Remove(f.Name())
+		return err
+	}
+
+	return os.Rename(f.Name(), s.path)
+}
+
+func (s *store) ResolveName(name model.Name) (model.Name, bool, error) {
+	// If a local model exists, do not allow alias shadowing (highest priority).
+	exists, err := localModelExists(name)
+	if err != nil {
+		return name, false, err
+	}
+	if exists {
+		return name, false, nil
+	}
+
+	key := normalizeAliasKey(name)
+
+	s.mu.RLock()
+	entry, exactMatch := s.entries[key]
+	var prefixMatch *aliasEntry
+	if !exactMatch {
+		// Try prefix matching - prefixEntries is sorted longest-first
+		nameStr := strings.ToLower(displayAliasName(name))
+		for i := range s.prefixEntries {
+			prefix := strings.ToLower(s.prefixEntries[i].Alias)
+			if strings.HasPrefix(nameStr, prefix) {
+				prefixMatch = &s.prefixEntries[i]
+				break // First match is longest due to sorting
+			}
+		}
+	}
+	s.mu.RUnlock()
+
+	if !exactMatch && prefixMatch == nil {
+		return name, false, nil
+	}
+
+	var current string
+	var visited map[string]struct{}
+
+	if exactMatch {
+		visited = map[string]struct{}{key: {}}
+		current = entry.Target
+	} else {
+		// For prefix match, use the target as-is
+		visited = map[string]struct{}{}
+		current = prefixMatch.Target
+	}
+
+	targetKey := normalizeAliasKeyString(current)
+
+	for {
+		targetName := model.ParseName(current)
+		if !targetName.IsValid() {
+			return name, false, fmt.Errorf("alias target %q is invalid", current)
+		}
+
+		if _, seen := visited[targetKey]; seen {
+			return name, false, errAliasCycle
+		}
+		visited[targetKey] = struct{}{}
+
+		s.mu.RLock()
+		next, ok := s.entries[targetKey]
+		s.mu.RUnlock()
+		if !ok {
+			return targetName, true, nil
+		}
+
+		current = next.Target
+		targetKey = normalizeAliasKeyString(current)
+	}
+}
+
+func (s *store) Set(alias, target model.Name, prefixMatching bool) error {
+	targetKey := normalizeAliasKey(target)
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	if prefixMatching {
+		// For prefix aliases, we skip cycle detection since prefix matching
+		// works differently and the target is a specific model
+		aliasStr := displayAliasName(alias)
+
+		// Remove any existing prefix entry with the same alias
+		for i, e := range s.prefixEntries {
+			if strings.EqualFold(e.Alias, aliasStr) {
+				s.prefixEntries = append(s.prefixEntries[:i], s.prefixEntries[i+1:]...)
+				break
+			}
+		}
+
+		s.prefixEntries = append(s.prefixEntries, aliasEntry{
+			Alias:          aliasStr,
+			Target:         displayAliasName(target),
+			PrefixMatching: true,
+		})
+		s.sortPrefixEntriesLocked()
+		return s.saveLocked()
+	}
+
+	aliasKey := normalizeAliasKey(alias)
+
+	if aliasKey == targetKey {
+		return fmt.Errorf("alias cannot point to itself")
+	}
+
+	visited := map[string]struct{}{aliasKey: {}}
+	currentKey := targetKey
+	for {
+		if _, seen := visited[currentKey]; seen {
+			return errAliasCycle
+		}
+		visited[currentKey] = struct{}{}
+
+		next, ok := s.entries[currentKey]
+		if !ok {
+			break
+		}
+		currentKey = normalizeAliasKeyString(next.Target)
+	}
+
+	s.entries[aliasKey] = aliasEntry{
+		Alias:  displayAliasName(alias),
+		Target: displayAliasName(target),
+	}
+
+	return s.saveLocked()
+}
+
+func (s *store) Delete(alias model.Name) (bool, error) {
+	aliasKey := normalizeAliasKey(alias)
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// Try exact match first
+	if _, ok := s.entries[aliasKey]; ok {
+		delete(s.entries, aliasKey)
+		return true, s.saveLocked()
+	}
+
+	// Try prefix entries
+	aliasStr := displayAliasName(alias)
+	for i, e := range s.prefixEntries {
+		if strings.EqualFold(e.Alias, aliasStr) {
+			s.prefixEntries = append(s.prefixEntries[:i], s.prefixEntries[i+1:]...)
+			return true, s.saveLocked()
+		}
+	}
+
+	return false, nil
+}
+
+// DeleteByString deletes an alias by its raw string value, useful for prefix
+// aliases that may not be valid model names.
+func (s *store) DeleteByString(alias string) (bool, error) {
+	alias = strings.TrimSpace(alias)
+	aliasLower := strings.ToLower(alias)
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// Try prefix entries first (since this is mainly for prefix aliases)
+	for i, e := range s.prefixEntries {
+		if strings.EqualFold(e.Alias, alias) {
+			s.prefixEntries = append(s.prefixEntries[:i], s.prefixEntries[i+1:]...)
+			return true, s.saveLocked()
+		}
+	}
+
+	// Also check exact entries by normalized key
+	if _, ok := s.entries[aliasLower]; ok {
+		delete(s.entries, aliasLower)
+		return true, s.saveLocked()
+	}
+
+	return false, nil
+}
+
+func (s *store) List() []aliasEntry {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+
+	entries := make([]aliasEntry, 0, len(s.entries)+len(s.prefixEntries))
+	for _, entry := range s.entries {
+		entries = append(entries, entry)
+	}
+	entries = append(entries, s.prefixEntries...)
+
+	sort.Slice(entries, func(i, j int) bool {
+		return strings.Compare(entries[i].Alias, entries[j].Alias) < 0
+	})
+	return entries
+}
+
+func normalizeAliasKey(name model.Name) string {
+	return strings.ToLower(displayAliasName(name))
+}
+
+func (s *store) sortPrefixEntriesLocked() {
+	sort.Slice(s.prefixEntries, func(i, j int) bool {
+		// Sort by length descending (longest prefix first)
+		return len(s.prefixEntries[i].Alias) > len(s.prefixEntries[j].Alias)
+	})
+}
+
+func normalizeAliasKeyString(value string) string {
+	n := model.ParseName(value)
+	if !n.IsValid() {
+		return strings.ToLower(strings.TrimSpace(value))
+	}
+	return normalizeAliasKey(n)
+}
+
+func displayAliasName(n model.Name) string {
+	display := n.DisplayShortest()
+	if strings.EqualFold(n.Tag, "latest") {
+		if idx := strings.LastIndex(display, ":"); idx != -1 {
+			return display[:idx]
+		}
+	}
+	return display
+}
+
+func localModelExists(name model.Name) (bool, error) {
+	manifests, err := manifest.Manifests(true)
+	if err != nil {
+		return false, err
+	}
+	needle := name.String()
+	for existing := range manifests {
+		if strings.EqualFold(existing.String(), needle) {
+			return true, nil
+		}
+	}
+	return false, nil
+}
+
+func serverConfigPath() string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return filepath.Join(".ollama", serverConfigFilename)
+	}
+	return filepath.Join(home, ".ollama", serverConfigFilename)
+}
+
+func (s *Server) aliasStore() (*store, error) {
+	s.aliasesOnce.Do(func() {
+		s.aliases, s.aliasesErr = createStore(serverConfigPath())
+	})
+
+	return s.aliases, s.aliasesErr
+}
+
+func (s *Server) resolveAlias(name model.Name) (model.Name, bool, error) {
+	store, err := s.aliasStore()
+	if err != nil {
+		return name, false, err
+	}
+
+	if store == nil {
+		return name, false, nil
+	}
+
+	return store.ResolveName(name)
+}

server/routes.go

@@ -22,6 +22,7 @@ import (
 	"os/signal"
 	"slices"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"syscall"
 	"time"
@@ -81,6 +82,9 @@ type Server struct {
 	addr          net.Addr
 	sched         *Scheduler
 	defaultNumCtx int
+	aliasesOnce   sync.Once
+	aliases       *store
+	aliasesErr    error
 }
 
 func init() {
@@ -191,9 +195,16 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
+	resolvedName, _, err := s.resolveAlias(name)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	name = resolvedName
+
 	// We cannot currently consolidate this into GetModel because all we'll
 	// induce infinite recursion given the current code structure.
-	name, err := getExistingName(name)
+	name, err = getExistingName(name)
 	if err != nil {
 		c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Model)})
 		return
@@ -1580,6 +1591,9 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.POST("/api/blobs/:digest", s.CreateBlobHandler)
 	r.HEAD("/api/blobs/:digest", s.HeadBlobHandler)
 	r.POST("/api/copy", s.CopyHandler)
+	r.GET("/api/experimental/aliases", s.ListAliasesHandler)
+	r.POST("/api/experimental/aliases", s.CreateAliasHandler)
+	r.DELETE("/api/experimental/aliases", s.DeleteAliasHandler)
 
 	// Inference
 	r.GET("/api/ps", s.PsHandler)
@@ -1950,13 +1964,20 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
-	name, err := getExistingName(name)
+	resolvedName, _, err := s.resolveAlias(name)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	name = resolvedName
+
+	name, err = getExistingName(name)
 	if err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": "model is required"})
 		return
 	}
 
-	m, err := GetModel(req.Model)
+	m, err := GetModel(name.String())
 	if err != nil {
 		switch {
 		case os.IsNotExist(err):

server/routes_aliases.go

@@ -0,0 +1,159 @@
+package server
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/ollama/ollama/types/model"
+)
+
+type aliasListResponse struct {
+	Aliases []aliasEntry `json:"aliases"`
+}
+
+type aliasDeleteRequest struct {
+	Alias string `json:"alias"`
+}
+
+func (s *Server) ListAliasesHandler(c *gin.Context) {
+	store, err := s.aliasStore()
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	var aliases []aliasEntry
+	if store != nil {
+		aliases = store.List()
+	}
+
+	c.JSON(http.StatusOK, aliasListResponse{Aliases: aliases})
+}
+
+func (s *Server) CreateAliasHandler(c *gin.Context) {
+	var req aliasEntry
+	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	} else if err != nil {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	req.Alias = strings.TrimSpace(req.Alias)
+	req.Target = strings.TrimSpace(req.Target)
+	if req.Alias == "" || req.Target == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "alias and target are required"})
+		return
+	}
+
+	// Target must always be a valid model name
+	targetName := model.ParseName(req.Target)
+	if !targetName.IsValid() {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("target %q is invalid", req.Target)})
+		return
+	}
+
+	var aliasName model.Name
+	if req.PrefixMatching {
+		// For prefix aliases, we still parse the alias to normalize it,
+		// but we allow any non-empty string since prefix patterns may not be valid model names
+		aliasName = model.ParseName(req.Alias)
+		// Even if not valid as a model name, we accept it for prefix matching
+	} else {
+		aliasName = model.ParseName(req.Alias)
+		if !aliasName.IsValid() {
+			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("alias %q is invalid", req.Alias)})
+			return
+		}
+
+		if normalizeAliasKey(aliasName) == normalizeAliasKey(targetName) {
+			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "alias cannot point to itself"})
+			return
+		}
+
+		exists, err := localModelExists(aliasName)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+		if exists {
+			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("alias %q conflicts with existing model", req.Alias)})
+			return
+		}
+	}
+
+	store, err := s.aliasStore()
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	if err := store.Set(aliasName, targetName, req.PrefixMatching); err != nil {
+		status := http.StatusInternalServerError
+		if errors.Is(err, errAliasCycle) {
+			status = http.StatusBadRequest
+		}
+		c.AbortWithStatusJSON(status, gin.H{"error": err.Error()})
+		return
+	}
+
+	resp := aliasEntry{
+		Alias:          displayAliasName(aliasName),
+		Target:         displayAliasName(targetName),
+		PrefixMatching: req.PrefixMatching,
+	}
+	if req.PrefixMatching && !aliasName.IsValid() {
+		// For prefix aliases that aren't valid model names, use the raw alias
+		resp.Alias = req.Alias
+	}
+	c.JSON(http.StatusOK, resp)
+}
+
+func (s *Server) DeleteAliasHandler(c *gin.Context) {
+	var req aliasDeleteRequest
+	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	} else if err != nil {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	req.Alias = strings.TrimSpace(req.Alias)
+	if req.Alias == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "alias is required"})
+		return
+	}
+
+	store, err := s.aliasStore()
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	aliasName := model.ParseName(req.Alias)
+	var deleted bool
+	if aliasName.IsValid() {
+		deleted, err = store.Delete(aliasName)
+	} else {
+		// For invalid model names (like prefix aliases), try deleting by raw string
+		deleted, err = store.DeleteByString(req.Alias)
+	}
+
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	if !deleted {
+		c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("alias %q not found", req.Alias)})
+		return
+	}
+
+	c.JSON(http.StatusOK, gin.H{"deleted": true})
+}

server/routes_aliases_test.go

@@ -0,0 +1,426 @@
+package server
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"path/filepath"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/types/model"
+)
+
+func TestAliasShadowingRejected(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	t.Setenv("HOME", t.TempDir())
+
+	s := Server{}
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
+		Model:      "shadowed-model",
+		RemoteHost: "example.com",
+		From:       "test",
+		Info: map[string]any{
+			"capabilities": []string{"completion"},
+		},
+		Stream: &stream,
+	})
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	w = createRequest(t, s.CreateAliasHandler, aliasEntry{Alias: "shadowed-model", Target: "other-model"})
+	if w.Code != http.StatusBadRequest {
+		t.Fatalf("expected status 400, got %d", w.Code)
+	}
+}
+
+func TestAliasResolvesForChatRemote(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	t.Setenv("HOME", t.TempDir())
+
+	var remoteModel string
+	rs := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var req api.ChatRequest
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			t.Fatal(err)
+		}
+		remoteModel = req.Model
+
+		w.Header().Set("Content-Type", "application/json")
+		resp := api.ChatResponse{
+			Model:      req.Model,
+			Done:       true,
+			DoneReason: "load",
+		}
+		if err := json.NewEncoder(w).Encode(&resp); err != nil {
+			t.Fatal(err)
+		}
+	}))
+	defer rs.Close()
+
+	p, err := url.Parse(rs.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	t.Setenv("OLLAMA_REMOTES", p.Hostname())
+
+	s := Server{}
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
+		Model:      "target-model",
+		RemoteHost: rs.URL,
+		From:       "test",
+		Info: map[string]any{
+			"capabilities": []string{"completion"},
+		},
+		Stream: &stream,
+	})
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	w = createRequest(t, s.CreateAliasHandler, aliasEntry{Alias: "alias-model", Target: "target-model"})
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	w = createRequest(t, s.ChatHandler, api.ChatRequest{
+		Model:    "alias-model",
+		Messages: []api.Message{{Role: "user", Content: "hi"}},
+		Stream:   &stream,
+	})
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	var resp api.ChatResponse
+	if err := json.NewDecoder(w.Body).Decode(&resp); err != nil {
+		t.Fatal(err)
+	}
+
+	if resp.Model != "alias-model" {
+		t.Fatalf("expected response model to be alias-model, got %q", resp.Model)
+	}
+
+	if remoteModel != "test" {
+		t.Fatalf("expected remote model to be 'test', got %q", remoteModel)
+	}
+}
+
+func TestPrefixAliasBasicMatching(t *testing.T) {
+	tmpDir := t.TempDir()
+	store, err := createStore(filepath.Join(tmpDir, "server.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Create a prefix alias: "myprefix-" -> "targetmodel"
+	targetName := model.ParseName("targetmodel")
+
+	// Set a prefix alias (using "myprefix-" as the pattern)
+	store.mu.Lock()
+	store.prefixEntries = append(store.prefixEntries, aliasEntry{
+		Alias:          "myprefix-",
+		Target:         "targetmodel",
+		PrefixMatching: true,
+	})
+	store.mu.Unlock()
+
+	// Test that "myprefix-foo" resolves to "targetmodel"
+	testName := model.ParseName("myprefix-foo")
+	resolved, wasResolved, err := store.ResolveName(testName)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !wasResolved {
+		t.Fatal("expected name to be resolved")
+	}
+	if resolved.DisplayShortest() != targetName.DisplayShortest() {
+		t.Fatalf("expected resolved name to be %q, got %q", targetName.DisplayShortest(), resolved.DisplayShortest())
+	}
+
+	// Test that "otherprefix-foo" does not resolve
+	otherName := model.ParseName("otherprefix-foo")
+	_, wasResolved, err = store.ResolveName(otherName)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if wasResolved {
+		t.Fatal("expected name not to be resolved")
+	}
+
+	// Test that exact alias takes precedence
+	exactAlias := model.ParseName("myprefix-exact")
+	exactTarget := model.ParseName("exacttarget")
+	if err := store.Set(exactAlias, exactTarget, false); err != nil {
+		t.Fatal(err)
+	}
+
+	resolved, wasResolved, err = store.ResolveName(exactAlias)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !wasResolved {
+		t.Fatal("expected name to be resolved")
+	}
+	if resolved.DisplayShortest() != exactTarget.DisplayShortest() {
+		t.Fatalf("expected resolved name to be %q (exact match), got %q", exactTarget.DisplayShortest(), resolved.DisplayShortest())
+	}
+}
+
+func TestPrefixAliasLongestMatchWins(t *testing.T) {
+	tmpDir := t.TempDir()
+	store, err := createStore(filepath.Join(tmpDir, "server.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Add two prefix aliases with overlapping patterns
+	store.mu.Lock()
+	store.prefixEntries = []aliasEntry{
+		{Alias: "abc-", Target: "short-target", PrefixMatching: true},
+		{Alias: "abc-def-", Target: "long-target", PrefixMatching: true},
+	}
+	store.sortPrefixEntriesLocked()
+	store.mu.Unlock()
+
+	// "abc-def-ghi" should match the longer prefix "abc-def-"
+	testName := model.ParseName("abc-def-ghi")
+	resolved, wasResolved, err := store.ResolveName(testName)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !wasResolved {
+		t.Fatal("expected name to be resolved")
+	}
+	expectedLongTarget := model.ParseName("long-target")
+	if resolved.DisplayShortest() != expectedLongTarget.DisplayShortest() {
+		t.Fatalf("expected resolved name to be %q (longest prefix match), got %q", expectedLongTarget.DisplayShortest(), resolved.DisplayShortest())
+	}
+
+	// "abc-xyz" should match the shorter prefix "abc-"
+	testName2 := model.ParseName("abc-xyz")
+	resolved, wasResolved, err = store.ResolveName(testName2)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !wasResolved {
+		t.Fatal("expected name to be resolved")
+	}
+	expectedShortTarget := model.ParseName("short-target")
+	if resolved.DisplayShortest() != expectedShortTarget.DisplayShortest() {
+		t.Fatalf("expected resolved name to be %q, got %q", expectedShortTarget.DisplayShortest(), resolved.DisplayShortest())
+	}
+}
+
+func TestPrefixAliasChain(t *testing.T) {
+	tmpDir := t.TempDir()
+	store, err := createStore(filepath.Join(tmpDir, "server.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Create a chain: prefix "test-" -> "intermediate" -> "final"
+	intermediate := model.ParseName("intermediate")
+	final := model.ParseName("final")
+
+	// Add prefix alias
+	store.mu.Lock()
+	store.prefixEntries = []aliasEntry{
+		{Alias: "test-", Target: "intermediate", PrefixMatching: true},
+	}
+	store.mu.Unlock()
+
+	// Add exact alias for the intermediate step
+	if err := store.Set(intermediate, final, false); err != nil {
+		t.Fatal(err)
+	}
+
+	// "test-foo" should resolve through the chain to "final"
+	testName := model.ParseName("test-foo")
+	resolved, wasResolved, err := store.ResolveName(testName)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !wasResolved {
+		t.Fatal("expected name to be resolved")
+	}
+	if resolved.DisplayShortest() != final.DisplayShortest() {
+		t.Fatalf("expected resolved name to be %q, got %q", final.DisplayShortest(), resolved.DisplayShortest())
+	}
+}
+
+func TestPrefixAliasCRUD(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	t.Setenv("HOME", t.TempDir())
+
+	s := Server{}
+
+	// Create a prefix alias via API
+	w := createRequest(t, s.CreateAliasHandler, aliasEntry{
+		Alias:          "myprefix-",
+		Target:         "llama2",
+		PrefixMatching: true,
+	})
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+
+	var createResp aliasEntry
+	if err := json.NewDecoder(w.Body).Decode(&createResp); err != nil {
+		t.Fatal(err)
+	}
+	if !createResp.PrefixMatching {
+		t.Fatal("expected prefix_matching to be true in response")
+	}
+
+	// List aliases and verify the prefix alias is included
+	w = createRequest(t, s.ListAliasesHandler, nil)
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	var listResp aliasListResponse
+	if err := json.NewDecoder(w.Body).Decode(&listResp); err != nil {
+		t.Fatal(err)
+	}
+
+	found := false
+	for _, a := range listResp.Aliases {
+		if a.PrefixMatching && a.Target == "llama2" {
+			found = true
+			break
+		}
+	}
+	if !found {
+		t.Fatal("expected to find prefix alias in list")
+	}
+
+	// Delete the prefix alias
+	w = createRequest(t, s.DeleteAliasHandler, aliasDeleteRequest{Alias: "myprefix-"})
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+
+	// Verify it's deleted
+	w = createRequest(t, s.ListAliasesHandler, nil)
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	if err := json.NewDecoder(w.Body).Decode(&listResp); err != nil {
+		t.Fatal(err)
+	}
+
+	for _, a := range listResp.Aliases {
+		if a.PrefixMatching {
+			t.Fatal("expected prefix alias to be deleted")
+		}
+	}
+}
+
+func TestPrefixAliasCaseInsensitive(t *testing.T) {
+	tmpDir := t.TempDir()
+	store, err := createStore(filepath.Join(tmpDir, "server.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Add a prefix alias with mixed case
+	store.mu.Lock()
+	store.prefixEntries = []aliasEntry{
+		{Alias: "MyPrefix-", Target: "targetmodel", PrefixMatching: true},
+	}
+	store.mu.Unlock()
+
+	// Test that matching is case-insensitive
+	testName := model.ParseName("myprefix-foo")
+	resolved, wasResolved, err := store.ResolveName(testName)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !wasResolved {
+		t.Fatal("expected name to be resolved (case-insensitive)")
+	}
+	expectedTarget := model.ParseName("targetmodel")
+	if resolved.DisplayShortest() != expectedTarget.DisplayShortest() {
+		t.Fatalf("expected resolved name to be %q, got %q", expectedTarget.DisplayShortest(), resolved.DisplayShortest())
+	}
+
+	// Test uppercase request
+	testName2 := model.ParseName("MYPREFIX-BAR")
+	_, wasResolved, err = store.ResolveName(testName2)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !wasResolved {
+		t.Fatal("expected name to be resolved (uppercase)")
+	}
+}
+
+func TestPrefixAliasLocalModelPrecedence(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	t.Setenv("HOME", t.TempDir())
+
+	s := Server{}
+
+	// Create a local model that would match a prefix alias
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
+		Model:      "myprefix-localmodel",
+		RemoteHost: "example.com",
+		From:       "test",
+		Info: map[string]any{
+			"capabilities": []string{"completion"},
+		},
+		Stream: &stream,
+	})
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+
+	// Create a prefix alias that would match the local model name
+	w = createRequest(t, s.CreateAliasHandler, aliasEntry{
+		Alias:          "myprefix-",
+		Target:         "someothermodel",
+		PrefixMatching: true,
+	})
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+
+	// Verify that resolving "myprefix-localmodel" returns the local model, not the alias target
+	store, err := s.aliasStore()
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	localModelName := model.ParseName("myprefix-localmodel")
+	resolved, wasResolved, err := store.ResolveName(localModelName)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if wasResolved {
+		t.Fatalf("expected local model to take precedence (wasResolved should be false), but got resolved to %q", resolved.DisplayShortest())
+	}
+	if resolved.DisplayShortest() != localModelName.DisplayShortest() {
+		t.Fatalf("expected resolved name to be local model %q, got %q", localModelName.DisplayShortest(), resolved.DisplayShortest())
+	}
+
+	// Also verify that a non-local model matching the prefix DOES resolve to the alias target
+	nonLocalName := model.ParseName("myprefix-nonexistent")
+	resolved, wasResolved, err = store.ResolveName(nonLocalName)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !wasResolved {
+		t.Fatal("expected non-local model to resolve via prefix alias")
+	}
+	expectedTarget := model.ParseName("someothermodel")
+	if resolved.DisplayShortest() != expectedTarget.DisplayShortest() {
+		t.Fatalf("expected resolved name to be %q, got %q", expectedTarget.DisplayShortest(), resolved.DisplayShortest())
+	}
+}

server/sched.go

@@ -417,9 +417,9 @@ func (s *Scheduler) load(req *LlmRequest, f *ggml.GGML, systemInfo ml.SystemInfo
 		numParallel = 1
 	}
 
-	// `mllama`, `qwen3vl`, and `qwen3vlmoe` are snowflakes and uses an encoder cache which cannot be used with num_parallel > 1
+	// Some architectures are not safe with num_parallel > 1.
 	// ref: https://github.com/ollama/ollama/issues/4165
-	if slices.Contains([]string{"mllama", "qwen3vl", "qwen3vlmoe"}, req.model.Config.ModelFamily) && numParallel != 1 {
+	if slices.Contains([]string{"mllama", "qwen3vl", "qwen3vlmoe", "qwen3next", "lfm2", "lfm2moe"}, req.model.Config.ModelFamily) && numParallel != 1 {
 		numParallel = 1
 		slog.Warn("model architecture does not currently support parallel requests", "architecture", req.model.Config.ModelFamily)
 	}