launch: add codex model metadata catalog

cmd/launch/codex.go

@@ -1,13 +1,20 @@
 package launch
 
 import (
+	"context"
+	"encoding/json"
 	"fmt"
+	"net/http"
 	"os"
 	"os/exec"
 	"path/filepath"
+	"slices"
+	"strconv"
 	"strings"
 
+	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/types/model"
 	"golang.org/x/mod/semver"
 )
 
@@ -32,7 +39,7 @@ func (c *Codex) Run(model string, args []string) error {
 		return err
 	}
 
-	if err := ensureCodexConfig(); err != nil {
+	if err := ensureCodexConfig(model); err != nil {
 		return fmt.Errorf("failed to configure codex: %w", err)
 	}
 
@@ -46,9 +53,9 @@ func (c *Codex) Run(model string, args []string) error {
 	return cmd.Run()
 }
 
-// ensureCodexConfig writes a [profiles.ollama-launch] section to ~/.codex/config.toml
-// with openai_base_url pointing to the local Ollama server.
-func ensureCodexConfig() error {
+// ensureCodexConfig writes a Codex profile and model catalog so Codex uses the
+// local Ollama server and has model metadata available.
+func ensureCodexConfig(modelName string) error {
 	home, err := os.UserHomeDir()
 	if err != nil {
 		return err
@@ -59,13 +66,18 @@ func ensureCodexConfig() error {
 		return err
 	}
 
+	catalogPath := filepath.Join(codexDir, "model.json")
+	if err := writeCodexModelCatalog(catalogPath, modelName); err != nil {
+		return err
+	}
+
 	configPath := filepath.Join(codexDir, "config.toml")
-	return writeCodexProfile(configPath)
+	return writeCodexProfile(configPath, catalogPath)
 }
 
 // writeCodexProfile ensures ~/.codex/config.toml has the ollama-launch profile
 // and model provider sections with the correct base URL.
-func writeCodexProfile(configPath string) error {
+func writeCodexProfile(configPath, catalogPath string) error {
 	baseURL := envconfig.Host().String() + "/v1/"
 
 	sections := []struct {
@@ -78,6 +90,7 @@ func writeCodexProfile(configPath string) error {
 				fmt.Sprintf("openai_base_url = %q", baseURL),
 				`forced_login_method = "api"`,
 				fmt.Sprintf("model_provider = %q", codexProfileName),
+				fmt.Sprintf("model_catalog_json = %q", catalogPath),
 			},
 		},
 		{
@@ -121,6 +134,110 @@ func writeCodexProfile(configPath string) error {
 	return os.WriteFile(configPath, []byte(text), 0o644)
 }
 
+func writeCodexModelCatalog(catalogPath, modelName string) error {
+	entry := buildCodexModelEntry(modelName)
+
+	catalog := map[string]any{
+		"models": []any{entry},
+	}
+
+	data, err := json.MarshalIndent(catalog, "", "  ")
+	if err != nil {
+		return err
+	}
+
+	return os.WriteFile(catalogPath, data, 0o644)
+}
+
+func buildCodexModelEntry(modelName string) map[string]any {
+	contextWindow := 0
+	hasVision := false
+	hasThinking := false
+	systemPrompt := ""
+
+	if l, ok := lookupCloudModelLimit(modelName); ok {
+		contextWindow = l.Context
+	}
+
+	client := api.NewClient(envconfig.Host(), http.DefaultClient)
+	resp, err := client.Show(context.Background(), &api.ShowRequest{Model: modelName})
+	if err == nil {
+		systemPrompt = resp.System
+		if slices.Contains(resp.Capabilities, model.CapabilityVision) {
+			hasVision = true
+		}
+		if slices.Contains(resp.Capabilities, model.CapabilityThinking) {
+			hasThinking = true
+		}
+
+		if !isCloudModelName(modelName) {
+			if n, ok := modelInfoContextLength(resp.ModelInfo); ok {
+				contextWindow = n
+			}
+			if resp.Details.Format != "safetensors" {
+				if ctxLen := envconfig.ContextLength(); ctxLen > 0 {
+					contextWindow = int(ctxLen)
+				}
+				if numCtx := parseNumCtx(resp.Parameters); numCtx > 0 {
+					contextWindow = numCtx
+				}
+			}
+		}
+	}
+
+	modalities := []string{"text"}
+	if hasVision {
+		modalities = append(modalities, "image")
+	}
+
+	reasoningLevels := []any{}
+	if hasThinking {
+		reasoningLevels = []any{
+			map[string]any{"effort": "low", "description": "Fast responses with lighter reasoning"},
+			map[string]any{"effort": "medium", "description": "Balances speed and reasoning depth"},
+			map[string]any{"effort": "high", "description": "Greater reasoning depth for complex problems"},
+		}
+	}
+
+	truncationMode := "bytes"
+	if isCloudModelName(modelName) {
+		truncationMode = "tokens"
+	}
+
+	return map[string]any{
+		"slug":                         modelName,
+		"display_name":                 modelName,
+		"context_window":               contextWindow,
+		"apply_patch_tool_type":        "function",
+		"shell_type":                   "default",
+		"visibility":                   "list",
+		"supported_in_api":             true,
+		"priority":                     0,
+		"truncation_policy":            map[string]any{"mode": truncationMode, "limit": 10000},
+		"input_modalities":             modalities,
+		"base_instructions":            systemPrompt,
+		"support_verbosity":            true,
+		"default_verbosity":            "low",
+		"supports_parallel_tool_calls": false,
+		"supports_reasoning_summaries": hasThinking,
+		"supported_reasoning_levels":   reasoningLevels,
+		"experimental_supported_tools": []any{},
+	}
+}
+
+func parseNumCtx(parameters string) int {
+	for _, line := range strings.Split(parameters, "\n") {
+		fields := strings.Fields(line)
+		if len(fields) == 2 && fields[0] == "num_ctx" {
+			if v, err := strconv.ParseFloat(fields[1], 64); err == nil {
+				return int(v)
+			}
+		}
+	}
+
+	return 0
+}
+
 func checkCodexVersion() error {
 	if _, err := exec.LookPath("codex"); err != nil {
 		return fmt.Errorf("codex is not installed, install with: npm install -g @openai/codex")

cmd/launch/codex_test.go

@@ -1,6 +1,10 @@
 package launch
 
 import (
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
 	"os"
 	"path/filepath"
 	"slices"
@@ -37,8 +41,9 @@ func TestWriteCodexProfile(t *testing.T) {
 	t.Run("creates new file when none exists", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -63,6 +68,9 @@ func TestWriteCodexProfile(t *testing.T) {
 		if !strings.Contains(content, `model_provider = "ollama-launch"`) {
 			t.Error("missing model_provider key")
 		}
+		if !strings.Contains(content, fmt.Sprintf("model_catalog_json = %q", catalogPath)) {
+			t.Error("missing model_catalog_json key")
+		}
 		if !strings.Contains(content, "[model_providers.ollama-launch]") {
 			t.Error("missing [model_providers.ollama-launch] section")
 		}
@@ -74,10 +82,11 @@ func TestWriteCodexProfile(t *testing.T) {
 	t.Run("appends profile to existing file without profile", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 		existing := "[some_other_section]\nkey = \"value\"\n"
 		os.WriteFile(configPath, []byte(existing), 0o644)
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -95,10 +104,11 @@ func TestWriteCodexProfile(t *testing.T) {
 	t.Run("replaces existing profile section", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 		existing := "[profiles.ollama-launch]\nopenai_base_url = \"http://old:1234/v1/\"\n\n[model_providers.ollama-launch]\nname = \"Ollama\"\nbase_url = \"http://old:1234/v1/\"\n"
 		os.WriteFile(configPath, []byte(existing), 0o644)
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -119,10 +129,11 @@ func TestWriteCodexProfile(t *testing.T) {
 	t.Run("replaces profile while preserving following sections", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 		existing := "[profiles.ollama-launch]\nopenai_base_url = \"http://old:1234/v1/\"\n[another_section]\nfoo = \"bar\"\n"
 		os.WriteFile(configPath, []byte(existing), 0o644)
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -143,10 +154,11 @@ func TestWriteCodexProfile(t *testing.T) {
 	t.Run("appends newline to file not ending with newline", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 		existing := "[other]\nkey = \"val\""
 		os.WriteFile(configPath, []byte(existing), 0o644)
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -166,8 +178,9 @@ func TestWriteCodexProfile(t *testing.T) {
 		t.Setenv("OLLAMA_HOST", "http://myhost:9999")
 		tmpDir := t.TempDir()
 		configPath := filepath.Join(tmpDir, "config.toml")
+		catalogPath := filepath.Join(tmpDir, "model.json")
 
-		if err := writeCodexProfile(configPath); err != nil {
+		if err := writeCodexProfile(configPath, catalogPath); err != nil {
 			t.Fatal(err)
 		}
 
@@ -185,7 +198,7 @@ func TestEnsureCodexConfig(t *testing.T) {
 		tmpDir := t.TempDir()
 		setTestHome(t, tmpDir)
 
-		if err := ensureCodexConfig(); err != nil {
+		if err := ensureCodexConfig("llama3.2"); err != nil {
 			t.Fatal(err)
 		}
 
@@ -202,16 +215,25 @@ func TestEnsureCodexConfig(t *testing.T) {
 		if !strings.Contains(content, "openai_base_url") {
 			t.Error("missing openai_base_url key")
 		}
+
+		catalogPath := filepath.Join(tmpDir, ".codex", "model.json")
+		data, err = os.ReadFile(catalogPath)
+		if err != nil {
+			t.Fatalf("model.json not created: %v", err)
+		}
+		if !strings.Contains(string(data), `"slug": "llama3.2"`) {
+			t.Error("missing model catalog entry for selected model")
+		}
 	})
 
 	t.Run("is idempotent", func(t *testing.T) {
 		tmpDir := t.TempDir()
 		setTestHome(t, tmpDir)
 
-		if err := ensureCodexConfig(); err != nil {
+		if err := ensureCodexConfig("llama3.2"); err != nil {
 			t.Fatal(err)
 		}
-		if err := ensureCodexConfig(); err != nil {
+		if err := ensureCodexConfig("llama3.2"); err != nil {
 			t.Fatal(err)
 		}
 
@@ -227,3 +249,204 @@ func TestEnsureCodexConfig(t *testing.T) {
 		}
 	})
 }
+
+func TestParseNumCtx(t *testing.T) {
+	tests := []struct {
+		name       string
+		parameters string
+		want       int
+	}{
+		{"num_ctx set", "num_ctx 8192", 8192},
+		{"num_ctx with other params", "temperature 0.7\nnum_ctx 4096\ntop_p 0.9", 4096},
+		{"no num_ctx", "temperature 0.7\ntop_p 0.9", 0},
+		{"empty string", "", 0},
+		{"malformed value", "num_ctx abc", 0},
+		{"float value", "num_ctx 8192.0", 8192},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := parseNumCtx(tt.parameters); got != tt.want {
+				t.Errorf("parseNumCtx(%q) = %d, want %d", tt.parameters, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestModelInfoContextLength(t *testing.T) {
+	tests := []struct {
+		name      string
+		modelInfo map[string]any
+		want      int
+	}{
+		{"float64 value", map[string]any{"qwen3_5_moe.context_length": float64(262144)}, 262144},
+		{"int value", map[string]any{"llama.context_length": 131072}, 131072},
+		{"no context_length key", map[string]any{"llama.embedding_length": float64(4096)}, 0},
+		{"empty map", map[string]any{}, 0},
+		{"nil map", nil, 0},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, _ := modelInfoContextLength(tt.modelInfo)
+			if got != tt.want {
+				t.Errorf("modelInfoContextLength() = %d, want %d", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestBuildCodexModelEntryContextWindow(t *testing.T) {
+	tests := []struct {
+		name          string
+		modelName     string
+		showResponse  string
+		envContextLen string
+		wantContext   int
+	}{
+		{
+			name:      "architectural context length as fallback",
+			modelName: "llama3.2",
+			showResponse: `{
+				"model_info": {"llama.context_length": 131072},
+				"details": {"format": "gguf"}
+			}`,
+			wantContext: 131072,
+		},
+		{
+			name:      "OLLAMA_CONTEXT_LENGTH overrides architectural",
+			modelName: "llama3.2",
+			showResponse: `{
+				"model_info": {"llama.context_length": 131072},
+				"details": {"format": "gguf"}
+			}`,
+			envContextLen: "64000",
+			wantContext:   64000,
+		},
+		{
+			name:      "num_ctx overrides OLLAMA_CONTEXT_LENGTH",
+			modelName: "llama3.2",
+			showResponse: `{
+				"model_info": {"llama.context_length": 131072},
+				"parameters": "num_ctx 8192",
+				"details": {"format": "gguf"}
+			}`,
+			envContextLen: "64000",
+			wantContext:   8192,
+		},
+		{
+			name:      "num_ctx overrides architectural",
+			modelName: "llama3.2",
+			showResponse: `{
+				"model_info": {"llama.context_length": 131072},
+				"parameters": "num_ctx 32768",
+				"details": {"format": "gguf"}
+			}`,
+			wantContext: 32768,
+		},
+		{
+			name:      "safetensors uses architectural context only",
+			modelName: "llama3.2",
+			showResponse: `{
+				"model_info": {"llama.context_length": 131072},
+				"parameters": "num_ctx 8192",
+				"details": {"format": "safetensors"}
+			}`,
+			envContextLen: "64000",
+			wantContext:   131072,
+		},
+		{
+			name:      "cloud model uses hardcoded limits",
+			modelName: "qwen3.5:cloud",
+			showResponse: `{
+				"model_info": {"qwen3_5_moe.context_length": 131072},
+				"details": {"format": "gguf"}
+			}`,
+			envContextLen: "64000",
+			wantContext:   262144,
+		},
+		{
+			name:      "vision and thinking capabilities",
+			modelName: "llama3.2",
+			showResponse: `{
+				"model_info": {"llama.context_length": 131072},
+				"details": {"format": "gguf"},
+				"capabilities": ["vision", "thinking"]
+			}`,
+			wantContext: 131072,
+		},
+		{
+			name:      "system prompt passed through",
+			modelName: "llama3.2",
+			showResponse: `{
+				"model_info": {"llama.context_length": 131072},
+				"details": {"format": "gguf"},
+				"system": "You are a helpful assistant."
+			}`,
+			wantContext: 131072,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				switch r.URL.Path {
+				case "/api/show":
+					fmt.Fprint(w, tt.showResponse)
+				default:
+					http.NotFound(w, r)
+				}
+			}))
+			defer srv.Close()
+			t.Setenv("OLLAMA_HOST", srv.URL)
+
+			if tt.envContextLen != "" {
+				t.Setenv("OLLAMA_CONTEXT_LENGTH", tt.envContextLen)
+			} else {
+				t.Setenv("OLLAMA_CONTEXT_LENGTH", "")
+			}
+
+			entry := buildCodexModelEntry(tt.modelName)
+
+			gotContext, _ := entry["context_window"].(int)
+			if gotContext != tt.wantContext {
+				t.Errorf("context_window = %d, want %d", gotContext, tt.wantContext)
+			}
+
+			if tt.name == "vision and thinking capabilities" {
+				modalities, _ := entry["input_modalities"].([]string)
+				if !slices.Contains(modalities, "image") {
+					t.Error("expected image in input_modalities")
+				}
+				levels, _ := entry["supported_reasoning_levels"].([]any)
+				if len(levels) == 0 {
+					t.Error("expected non-empty supported_reasoning_levels")
+				}
+			}
+
+			if tt.name == "system prompt passed through" {
+				if got, _ := entry["base_instructions"].(string); got != "You are a helpful assistant." {
+					t.Errorf("base_instructions = %q, want %q", got, "You are a helpful assistant.")
+				}
+			}
+
+			if tt.name == "cloud model uses hardcoded limits" {
+				truncationPolicy, _ := entry["truncation_policy"].(map[string]any)
+				if mode, _ := truncationPolicy["mode"].(string); mode != "tokens" {
+					t.Errorf("truncation_policy mode = %q, want %q", mode, "tokens")
+				}
+			}
+
+			requiredKeys := []string{"slug", "display_name", "apply_patch_tool_type", "shell_type"}
+			for _, key := range requiredKeys {
+				if _, ok := entry[key]; !ok {
+					t.Errorf("missing required key %q", key)
+				}
+			}
+
+			if _, err := json.Marshal(entry); err != nil {
+				t.Errorf("entry is not JSON serializable: %v", err)
+			}
+		})
+	}
+}