@lasso-ai/cli 1.2.4 → 1.2.6

package/ARCHITECTURE.md

@@ -85,6 +85,13 @@ Responsibilities, and _only_ these:
   `_debugSource` fiber data.
 - Take a screenshot of the selection (`html2canvas`) — visual context measurably improves
   edit quality even when the AI has the source.
+- Capture the page's runtime health: uncaught `window` errors, unhandled promise
+  rejections, and the host app's `console.error` output. Listeners are installed when
+  the overlay script evaluates (before the app's own modules), so failures during
+  initial load — a broken import, a missing named export — are recorded as well. The
+  latest five unique entries travel with every request (`context.runtimeErrors` for
+  select/drag, `pageContext.runtimeErrors` for one-shot) and are cleared on apply so
+  they always describe the code that is running now.
 - Show the inline prompt box anchored to the selection (never a blocking `prompt()`).
 - Render the diff/accept/undo UI once the CLI streams a proposed change back.
 - Render concurrent Agent Tasks with In progress/Completed tabs and task-level review
@@ -118,8 +125,8 @@ A Node process that wraps the user's existing dev server. Entry: `src/cli/index.
    never touched on disk.
 3. Host the WebSocket bridge (`src/cli/bridge.ts`) the browser overlay connects to.
 4. On a selection + instruction: read the referenced source file(s), assemble context
-   (source snippet, imports, relevant CSS/Tailwind config, screenshot, instruction), and
-   hand it to the coding agent (§6).
+   (source snippet, imports, relevant CSS/Tailwind config, screenshot, captured
+   console/runtime errors, instruction), and hand it to the coding agent (§6).
 5. Hold the returned diff in memory and stream it to the browser for preview — never write
    on receipt.
 6. On accept: snapshot the file's prior content (for undo), then apply the diff.
@@ -163,7 +170,7 @@ Lasso ships its own default agent, but the CLI's job is to assemble _context_, n
 married to one model.
 
 ```
-context (source + screenshot + instruction)
+context (source + screenshot + console errors + instruction)
         │
         ▼
   ┌───────────────┐

package/CHANGELOG.md

@@ -7,13 +7,56 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Fixed
+
+- **One-shot generation no longer demands a provider API key when a local agent is
+  selected.** Picking Claude Code, Codex, OpenCode, or Cursor for a one-shot request
+  used to bail out with *"Please configure an AI provider API key or model in settings
+  to enable code generation."* Those agents authenticate themselves, so the request now
+  runs through the local CLI first. If the local CLI cannot run (not installed, not
+  signed in, out of credits) and a Lasso key is available, the request falls back to the
+  Lasso Agent Gateway with a hosted provider/model it understands (`claude-code`/`cursor`
+  → Anthropic, `codex` → OpenAI, `opencode` by model prefix). Without a key, the error
+  now names the CLI command and `lasso auth login` instead of pointing at settings.
+- **Local agents no longer hang on stdin.** `codex exec` and `opencode run` read stdin
+  until EOF, so every local-agent request stalled until the 5-minute timeout. Stdin is
+  closed as soon as the process spawns (none of these command lines are interactive).
+- **Ollama requests stay local.** One-shot requests for Ollama are no longer routed to
+  the gateway, which cannot reach the user's `localhost`.
+
 ### Changed
 
+- **`lasso init` registers against the Lasso API, not the realtime server.**
+  Project registration (`init`, `register`, `create`) now posts to
+  `serverUrlFrom()` — `https://api.lasso.byorello.space` by default — because the
+  API owns the workspace's project records and stores the local `.lasso` domain.
+  The realtime/collab server (`…/api/v1/collab/projects/register`) is only tried
+  when the API cannot be reached. Override with `LASSO_API_URL` /
+  `LASSO_SERVER_URL`.
 - **One-shot mode as sidebar panel**: Changed one-shot mode from a centered modal to a right-side sidebar panel that only appears during one-shot mode. When users exit one-shot mode, they return to the normal edit mode with just the toolbar. This provides a cleaner, more integrated experience similar to modern AI editors.
 - **Mini toolbar positioning**: Adjusted the mini toolbar position to `left: calc(50% + 140px)` (reduced from 280px) to reduce overlap with the main toolbar. Added smooth transition and centered positioning when the main toolbar is dismissed. Mini toolbar is now hidden when one-shot sidebar is open to avoid UI clutter.
 
 ### Added
 
+- **Console-aware agent context**: the overlay now captures the page's runtime
+  health — uncaught `window` errors, unhandled promise rejections, and anything
+  the host app logs through `console.error` — starting from the moment the overlay
+  script loads, so failures during initial page load (for example
+  `SyntaxError: The requested module … does not provide an export named …` from a
+  broken import) are recorded too. The latest five unique entries (1200 chars each)
+  are attached to every request:
+  - **Select / component flow**: `edit`, `ask`, and drag requests carry
+    `context.runtimeErrors`, which the agent prompt renders as an explicit
+    `LIVE CONSOLE/RUNTIME ERRORS` block alongside the element and source context.
+  - **One-shot flow**: `pageContext.runtimeErrors` travels with the ⌘K request and
+    renders as a `Console/Runtime Errors` line in the live page context — local
+    plan/conversation prompts and the hosted agent gateway alike.
+- **Runtime errors are part of the task**: agent instructions treat listed console
+  errors as work to fix, and require verifying named package exports (via
+  `run_command` over `node_modules` or the package docs) before importing them, so a
+  guessed export name cannot kill the page with a SyntaxError.
+- **Error buffer resets on apply**: applying a change clears the captured runtime
+  errors, so the next prompt only reflects errors produced by the new code.
 - **Multiple access points for one-shot mode**: Added a Command button to the toolbar (⌘ icon) for opening the one-shot sidebar, in addition to the existing ⌘K keyboard shortcut. The sidebar includes a model selector dropdown allowing users to choose their AI model before submitting a one-shot request.
 - **Model selector in command sidebar**: The one-shot sidebar now displays the currently selected model with a dropdown to switch between available models. The selector is organized by provider (Anthropic, OpenAI, Google, Local Agents) for easy navigation.
 - **Cursor CLI support**: Added Cursor as a local agent option alongside Claude Code, Codex, and OpenCode. Users can now select Cursor models in the model dropdown if Cursor is installed on their system. Cursor uses similar output parsing to Claude Code with progress tracking and error handling.
@@ -130,6 +173,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   name, version, workspaceId }` so the overlay can join the right (workspace-
   gated) session. The bridge only forwards a config it successfully authenticated.
 
+### Fixed
+
+- **Useless `Project registration failed (400)` from `lasso init`**: the CLI only
+  read `body.message`, but the registration endpoint answers with `{ error }`, so
+  every server-side validation message was dropped and collapsed into the generic
+  status line. The server text is surfaced now, e.g.
+  `"domain" is not allowed (HTTP 400)`.
+- **Selected element missing from one-shot page context**: `capturePageContext()`
+  read `state.selectedElement`, which does not exist on the overlay state (the
+  property is `state.selected`), so the inspected element never reached the page
+  context. The same mistake failed `typecheck:overlay`.
+
 ## [0.1.0] - 2026-09-21
 
 ### Added

package/README.md

@@ -66,7 +66,7 @@ Lasso works differently.
 | | Lasso |
 |---|---|
 | 🎯 **Point at the UI** | Select exactly what you want to change |
-| 🧠 **AI understands context** | The selected component is resolved back to your source |
+| 🧠 **AI understands context** | The selected component is resolved back to your source, plus the live page state (DOM summary, screenshot, console/runtime errors) |
 | 📝 **Real code changes** | Lasso modifies your actual source files |
 | 👀 **Preview first** | Every change comes back as a diff |
 | 🔒 **You stay in control** | Nothing is written until you explicitly accept |
@@ -97,6 +97,24 @@ Lasso now supports two equally powerful ways to build:
 These modes work together naturally: use one-shot to generate large features, then
 visual mode to refine specific parts.
 
+### Console-aware context
+
+Both modes receive the page's runtime health alongside the DOM and screenshot
+context. The overlay captures:
+
+- uncaught `window` errors and unhandled promise rejections
+- anything your app logs through `console.error`
+
+Capture starts with the overlay script itself, so failures during the initial page
+load (a broken import, a missing named export) are recorded too. The latest five
+unique entries ride along with every request — as `context.runtimeErrors` for
+select/drag prompts and as `pageContext.runtimeErrors` for ⌘K one-shot prompts —
+and the agent is instructed to treat listed errors as part of the task instead of
+working around them.
+
+The buffer is cleared when you apply a change, so the next prompt only reflects
+errors produced by the new code.
+
 ### Visual selection
 
 Turn on Lasso Mode and interact with your application normally.
@@ -156,7 +174,8 @@ or:
 
 > "Add a loading state below this button."
 
-Lasso assembles the relevant source context and sends it to your configured coding agent.
+Lasso assembles the relevant source context — source files, screenshot, and the
+page's console/runtime errors — and sends it to your configured coding agent.
 
 ### Diff-first workflow
 
@@ -496,6 +515,7 @@ Lasso has three main pieces.
 │                      │
 │  Select UI element   │
 │  Capture screenshot  │
+│  Console errors      │
 │  Presence, locks,    │
 │  comments, voice     │
 └──────────┬───────────┘
@@ -544,7 +564,7 @@ The CLI:
 2. Resolves the selected component
 3. Reads the relevant source
 4. Collects imports and surrounding context
-5. Captures the selected UI
+5. Captures the selected UI and the page's console/runtime errors
 6. Sends the assembled context to your configured agent
 7. Receives a proposed change
 8. Shows you the diff

package/dist/cli/agent.d.ts

@@ -3,7 +3,7 @@ export type SourceChange = {
     oldString: string;
     newString: string;
 };
-type AgentInput = {
+export type AgentInput = {
     taskId?: string;
     instruction: string;
     model: string;
@@ -97,6 +97,7 @@ export type AgentStepAction = {
     changes: SourceChange[];
 };
 export declare function parseAgentAction(text: string): AgentStepAction;
+export declare function runLocalAgentPrompt(cwd: string, config: AgentConfig, prompt: string, signal?: AbortSignal, onProgress?: AgentProgress, taskId?: string, onPrompt?: AgentPromptHandler): Promise<string>;
 export declare function proposeChanges(cwd: string, input: AgentInput, config: AgentConfig, signal?: AbortSignal, onProgress?: AgentProgress, onPrompt?: AgentPromptHandler): Promise<{
     summary: string;
     packages?: string[];
@@ -104,4 +105,3 @@ export declare function proposeChanges(cwd: string, input: AgentInput, config: A
 }>;
 export declare function answerQuestion(cwd: string, input: AgentAnswer, config: AgentConfig, signal?: AbortSignal, onProgress?: AgentProgress, onPrompt?: AgentPromptHandler): Promise<string>;
 export declare function generateCommitMessage(cwd: string, status: string[], config: AgentConfig, signal?: AbortSignal, onProgress?: AgentProgress): Promise<string>;
-export {};

package/dist/cli/agent.js

@@ -10,6 +10,7 @@ exports.executeWebSearch = executeWebSearch;
 exports.executeAgentCommand = executeAgentCommand;
 exports.detectLocalAgents = detectLocalAgents;
 exports.parseAgentAction = parseAgentAction;
+exports.runLocalAgentPrompt = runLocalAgentPrompt;
 exports.proposeChanges = proposeChanges;
 exports.answerQuestion = answerQuestion;
 exports.generateCommitMessage = generateCommitMessage;
@@ -755,6 +756,12 @@ function localCommand(provider, model, prompt) {
 function runLocalCommand(cwd, command, args, provider, signal, onProgress, taskId, onPrompt) {
     return new Promise((resolve, reject) => {
         const child = (0, node_child_process_1.spawn)(command, args, { cwd, stdio: ["pipe", "pipe", "pipe"] });
+        // Headless invocations: codex/opencode/cursor read stdin until EOF and otherwise hang
+        // until our timeout, so close stdin right away. Prompt answers can no longer be piped
+        // back in, which is fine — none of these command lines are interactive. Swallow the
+        // EPIPE that a late write would raise so it cannot crash the CLI.
+        child.stdin?.on("error", () => { });
+        child.stdin?.end();
         const label = provider === "opencode" ? "OpenCode" : provider === "codex" ? "Codex" : provider === "cursor" ? "Cursor" : "Claude Code";
         let stdout = "";
         let stderr = "";
@@ -882,6 +889,16 @@ async function proposeWithLocalAgent(cwd, instruction, context, config, signal,
         throw new Error(localAgentError(command, result.stderr, result.exitCode));
     return extractLocalAgentProposal(result.stdout, config.provider);
 }
+// Runs a local coding-agent CLI (claude/codex/opencode/cursor) and returns its final text.
+// The CLI authenticates itself (its own login), so no provider API key is required.
+async function runLocalAgentPrompt(cwd, config, prompt, signal, onProgress, taskId, onPrompt) {
+    const provider = config.provider;
+    const local = localCommand(provider, config.model, prompt);
+    const result = await runLocalCommand(cwd, local.command, local.args, provider, signal, onProgress, taskId, onPrompt);
+    if (result.exitCode !== 0)
+        throw new Error(localAgentError(local.command, result.stderr, result.exitCode));
+    return extractLocalAgentText(result.stdout, provider);
+}
 async function proposeChanges(cwd, input, config, signal, onProgress, onPrompt) {
     const context = await contextFor(cwd, input.element);
     const runtimeErrors = input.context?.runtimeErrors || [];
@@ -1176,11 +1193,7 @@ async function answerQuestion(cwd, input, config, signal, onProgress, onPrompt)
     const history = input.messages?.map((message) => `${message.role}: ${message.content}`).join("\n") || "None";
     const prompt = `Answer the user's question conversationally and directly. Do not propose file changes and do not return JSON. If the question is about the selected UI, use the selection and source context below.\n\nUser question:\n${input.question}\n\nSelected element:\n${JSON.stringify(input.element, null, 2)}\n\nVisual context:\n${JSON.stringify({ ...input.context, screenshots: undefined }, null, 2)}\n\nConversation:\n${history}\n\nRelevant source context:\n${context || "No matching source context was found."}`;
     if (config.provider === "claude-code" || config.provider === "codex" || config.provider === "opencode" || config.provider === "cursor") {
-        const local = localCommand(config.provider, config.model, prompt);
-        const result = await runLocalCommand(cwd, local.command, local.args, config.provider, signal, onProgress, input.taskId, onPrompt);
-        if (result.exitCode !== 0)
-            throw new Error(localAgentError(local.command, result.stderr, result.exitCode));
-        return extractLocalAgentText(result.stdout, config.provider).trim();
+        return (await runLocalAgentPrompt(cwd, config, prompt, signal, onProgress, input.taskId, onPrompt)).trim();
     }
     const callServerGatewayAnswer = async (gatewayKey) => {
         let serverUrl = config.serverUrl || process.env.LASSO_SERVER_URL || process.env.NEXT_PUBLIC_LASSO_SERVER_URL || "https://api.lasso.byorello.space";

package/dist/cli/bridge.js

@@ -556,7 +556,7 @@ function startBridge(cwd = process.cwd(), collabConfig = null, bridgePort = expo
                 taskControllers.set(msg.taskId, controller);
                 let selectedConfig;
                 if (localProvider) {
-                    selectedConfig = { provider: cliProvider, model: msg.model };
+                    selectedConfig = { provider: cliProvider, model: msg.model, lassoKey: resolvedLassoKey, serverUrl: collabConfig?.serverUrl || (0, auth_1.serverUrlFrom)(fileEnv) };
                 }
                 else {
                     const requestedProvider = (msg.provider || agentConfig.provider);
@@ -608,7 +608,7 @@ function startBridge(cwd = process.cwd(), collabConfig = null, bridgePort = expo
                 }
                 let selectedConfig;
                 if (localProvider) {
-                    selectedConfig = { provider: cliProvider, model: msg.model };
+                    selectedConfig = { provider: cliProvider, model: msg.model, lassoKey: resolvedLassoKey, serverUrl: collabConfig?.serverUrl || (0, auth_1.serverUrlFrom)(fileEnv) };
                 }
                 else {
                     const requestedProvider = (msg.provider || agentConfig.provider);
@@ -765,21 +765,23 @@ function startBridge(cwd = process.cwd(), collabConfig = null, bridgePort = expo
                 const provider = (msg.provider || agentConfig?.provider || "google");
                 const model = msg.model || agentConfig?.model || (provider === "nvidia" ? "nvidia/llama-3.1-nemotron-70b-instruct" : provider === "google" ? "gemini-2.5-flash" : "claude-3-7-sonnet");
                 const localKey = localProviderKeys[provider];
+                const serverUrl = collabConfig?.serverUrl || (0, auth_1.serverUrlFrom)(fileEnv);
+                const targetApiUrl = (serverUrl.includes("collab.lasso.byorello.space") ? "https://api.lasso.byorello.space" : serverUrl).replace(/\/api\/v1\/?$/, "");
                 const selectedConfig = localProvider
-                    ? { provider: cliProvider, model: msg.model || "claude-code:sonnet" }
+                    ? { provider: cliProvider, model: msg.model || "claude-code:sonnet", lassoKey: resolvedLassoKey, serverUrl: targetApiUrl }
                     : {
                         ...(agentConfig || {}),
                         provider,
                         model,
                         apiKey: localKey || apiKey,
+                        lassoKey: resolvedLassoKey,
+                        serverUrl: targetApiUrl,
                         baseUrl: provider === "nvidia" ? "https://integrate.api.nvidia.com/v1" : provider === "ollama" ? `${(process.env.OLLAMA_BASE_URL || fileEnv.OLLAMA_BASE_URL || "http://localhost:11434/api").replace(/\/api\/?$/, "")}/v1` : undefined,
                     };
                 const controller = new AbortController();
                 taskControllers.set(taskId, controller);
                 socket.send(JSON.stringify({ type: "agent_status", taskId, status: "thinking", message: "Planning your request..." }));
                 const { runOneShotAgent } = await import("./oneshot.js");
-                const serverUrl = collabConfig?.serverUrl || (0, auth_1.serverUrlFrom)(fileEnv);
-                const targetApiUrl = (serverUrl.includes("collab.lasso.byorello.space") ? "https://api.lasso.byorello.space" : serverUrl).replace(/\/api\/v1\/?$/, "");
                 runOneShotAgent(cwd, msg.prompt, msg.scope || "project", selectedConfig, controller.signal, (status, message, detail) => {
                     if (!controller.signal.aborted && socket.readyState === socket.OPEN) {
                         socket.send(JSON.stringify({ type: "agent_status", taskId, status, message, detail }));

package/dist/cli/oneshot.js

@@ -9,7 +9,37 @@ const node_fs_1 = __importDefault(require("node:fs"));
 const node_path_1 = __importDefault(require("node:path"));
 const node_child_process_1 = require("node:child_process");
 const node_util_1 = require("node:util");
+const agent_1 = require("./agent");
 const execAsync = (0, node_util_1.promisify)(node_child_process_1.exec);
+const CLI_PROVIDERS = new Set(["claude-code", "codex", "opencode", "cursor"]);
+function isCliProvider(provider) {
+    return Boolean(provider && CLI_PROVIDERS.has(provider));
+}
+function cliLabel(provider) {
+    return provider === "claude-code" ? "Claude Code" : provider === "opencode" ? "OpenCode" : provider === "cursor" ? "Cursor" : provider === "codex" ? "Codex" : "CLI agent";
+}
+function cliCommand(provider) {
+    return provider === "claude-code" ? "claude" : provider || "the agent CLI";
+}
+// The Lasso Agent Gateway only understands hosted providers, so a CLI selection is
+// normalized to the closest hosted provider/model when we have to fall back to it.
+const GATEWAY_MODELS = {
+    anthropic: "claude-sonnet-4-5-20250929",
+    openai: "gpt-4.1",
+    google: "gemini-2.5-flash",
+};
+function gatewayTargetForCli(provider, model) {
+    const bare = model && model.includes(":") ? model.slice(model.indexOf(":") + 1) : model || "";
+    if (provider === "codex")
+        return { provider: "openai", model: GATEWAY_MODELS.openai };
+    if (provider === "cursor")
+        return bare.includes("gpt") ? { provider: "openai", model: GATEWAY_MODELS.openai } : { provider: "anthropic", model: GATEWAY_MODELS.anthropic };
+    if (bare.startsWith("openai/") || /^gpt/i.test(bare))
+        return { provider: "openai", model: GATEWAY_MODELS.openai };
+    if (bare.startsWith("google/") || /^gemini/i.test(bare))
+        return { provider: "google", model: GATEWAY_MODELS.google };
+    return { provider: "anthropic", model: GATEWAY_MODELS.anthropic };
+}
 function isConversationalPrompt(prompt) {
     const p = prompt.trim().toLowerCase().replace(/[!?.,;:]+$/, "");
     const greetings = [
@@ -39,7 +69,7 @@ function runtimeErrorObservation(pageContext) {
         return "- Console/Runtime Errors: None";
     return `- Console/Runtime Errors (the page is currently throwing these):\n${errors.map((e) => `  ! ${e}`).join("\n")}`;
 }
-async function generateConversationalReply(prompt, framework, config, signal, messages, pageContext) {
+async function generateConversationalReply(cwd, prompt, framework, config, signal, messages, pageContext, onProgress, taskId, onPrompt) {
     const history = messages?.map((m) => `${m.role === "user" ? "User" : "Assistant"}: ${m.content}`).join("\n\n") || "";
     let pageObservation = "";
     if (pageContext) {
@@ -68,13 +98,19 @@ Respond conversationally, concisely, and helpfully.
 ${pageContext ? "You have live access to the browser page context. When the user asks about the page, design, layout, or components, reference the actual headings, sections, buttons, and text rendered on screen. Provide concrete, expert design and UX feedback. Never claim the page has no content or components when rendered headings and sections are present." : ""}
 Do not output code changes, file patches, or JSON schemas — just talk to the developer naturally and offer assistance.`;
     const userPrompt = `${history ? `Conversation History:\n${history}\n\n` : ""}${pageObservation}\nUser message: ${prompt}`;
+    // CLI agents (Claude Code, Codex, OpenCode, Cursor) authenticate themselves, so no
+    // provider API key is needed. Failures are thrown so the caller can fall back.
+    if (isCliProvider(config.provider)) {
+        const text = await (0, agent_1.runLocalAgentPrompt)(cwd, config, `${systemPrompt}\n\n${userPrompt}`, signal, onProgress, taskId, onPrompt);
+        return text.trim() || "Hello! I'm Lasso, your AI pair programmer. You can ask me questions about your project or tell me what to build, modify, or fix.";
+    }
     if (!config.apiKey && config.provider !== "ollama") {
         if (pageContext && (prompt.toLowerCase().includes("landing page") || prompt.toLowerCase().includes("page"))) {
             const headings = (pageContext.domSummary?.headings || []).map((h) => `"${h.text}"`).join(", ");
             const sections = (pageContext.domSummary?.sections || []).map((s) => s.name).join(", ");
-            return `Looking at your current page ("${pageContext.title}" on ${pageContext.route}):\n\n- **Rendered Sections:** ${sections || "Standard container"}\n- **Key Headings:** ${headings || "None"}\n- **CTAs:** ${(pageContext.domSummary?.buttons || []).join(", ") || "None"}\n\nTo build or refine components, configure your AI API key in settings or ask me what to modify!`;
+            return `Looking at your current page ("${pageContext.title}" on ${pageContext.route}):\n\n- **Rendered Sections:** ${sections || "Standard container"}\n- **Key Headings:** ${headings || "None"}\n- **CTAs:** ${(pageContext.domSummary?.buttons || []).join(", ") || "None"}\n\nTo build or refine components, run \`lasso auth login\` or add an AI provider API key in settings, then ask me what to modify!`;
         }
-        return "Hello! I'm Lasso, your AI pair programmer. You can ask me questions about your project or tell me what to build, modify, or fix.";
+        return "Hello! I'm Lasso, your AI pair programmer. You can ask me questions about your project or tell me what to build, modify, or fix. Run `lasso auth login` (or add a provider API key in settings) to unlock AI replies.";
     }
     try {
         const model = config.model || (config.provider === "google" ? "gemini-2.5-flash" : config.provider === "openai" ? "gpt-4.1-mini" : config.provider === "ollama" ? "llama3.2" : config.provider === "nvidia" ? "nvidia/llama-3.1-nemotron-70b-instruct" : "claude-3-7-sonnet-latest");
@@ -179,21 +215,20 @@ async function runOneShotAgent(cwd, prompt, scope, config, signal, onProgress, o
             rawServerUrl = rawServerUrl.replace("collab.lasso.byorello.space", "api.lasso.byorello.space");
         }
         const targetServerUrl = rawServerUrl.replace(/\/$/, "").replace(/\/api\/v1$/, "");
-        const isCliProvider = config.provider === "claude-code" ||
-            config.provider === "codex" ||
-            config.provider === "opencode" ||
-            config.provider === "cursor";
+        const useCliAgent = isCliProvider(config.provider);
+        const gatewayKey = apiKey || config.lassoKey || (config.apiKey && (config.apiKey.startsWith("lss_live_") || config.apiKey.startsWith("lss_")) ? config.apiKey : "") || process.env.LASSO_API_KEY || "";
         const hasLocalProviderKey = Boolean(config.apiKey &&
             config.apiKey !== "ollama" &&
             !config.apiKey.startsWith("lss_live_") &&
             !config.apiKey.startsWith("lss_"));
         // If the user does NOT have a local provider key in .env, route to Lasso Agent Gateway on server.
         // When they DO have a local key in .env, it runs locally to save server resources!
-        const shouldRouteToServer = !isCliProvider && !hasLocalProviderKey && Boolean(targetServerUrl);
-        if (shouldRouteToServer) {
+        // Ollama always runs locally (the server cannot reach the user's localhost).
+        const shouldRouteToServer = !useCliAgent && config.provider !== "ollama" && !hasLocalProviderKey && Boolean(targetServerUrl);
+        const callGateway = async (target) => {
             try {
                 if (!isConversational && !pageContext) {
-                    onProgress("thinking", "Connecting to Lasso Agent Gateway...", `Model: ${config.model || "claude-3-7-sonnet"}`);
+                    onProgress("thinking", "Connecting to Lasso Agent Gateway...", `Model: ${target.model || "claude-3-7-sonnet"}`);
                 }
                 const endpoint = `${targetServerUrl.replace(/\/$/, "")}/api/v1/agent/session`;
                 const headers = {
@@ -235,8 +270,8 @@ async function runOneShotAgent(cwd, prompt, scope, config, signal, onProgress, o
                         prompt,
                         messages,
                         model: {
-                            id: config.model,
-                            provider: config.provider,
+                            id: target.model,
+                            provider: target.provider,
                         },
                         context: {
                             framework,
@@ -315,10 +350,100 @@ async function runOneShotAgent(cwd, prompt, scope, config, signal, onProgress, o
                     error: serverErr instanceof Error ? serverErr.message : "Failed to connect to Lasso Agent Gateway.",
                 };
             }
+            return { ok: false, error: "The agent session returned no result." };
+        };
+        const gatewayFallback = async (reason) => {
+            if (!gatewayKey || !targetServerUrl) {
+                return {
+                    ok: false,
+                    error: `${reason} — run \`${cliCommand(config.provider)}\` once to authenticate it, or run \`lasso auth login\` to use the Lasso gateway.`,
+                };
+            }
+            onProgress("thinking", `${cliLabel(config.provider)} is unavailable; using the Lasso Agent Gateway...`, reason);
+            const target = gatewayTargetForCli(config.provider, config.model);
+            const result = await callGateway({ provider: target.provider, model: target.model });
+            if (!result.ok && result.error) {
+                return { ok: false, error: `${reason} — gateway fallback failed: ${result.error}` };
+            }
+            return result;
+        };
+        // CLI agents (Claude Code, Codex, OpenCode, Cursor) authenticate themselves, so they run
+        // locally first instead of demanding a provider API key from settings.
+        if (useCliAgent && !isConversational) {
+            onProgress("thinking", `Running ${cliLabel(config.provider)}...`, `Local CLI agent · ${config.model || ""}`);
+            try {
+                const selectedElement = pageContext?.selectedElement;
+                const input = {
+                    taskId,
+                    instruction: prompt,
+                    model: config.model || "",
+                    messages,
+                    context: {
+                        runtimeErrors: pageContext?.runtimeErrors,
+                        screenshots: pageContext?.screenshot ? { full: pageContext.screenshot } : undefined,
+                        viewport: pageContext?.viewport,
+                    },
+                    element: {
+                        tag: selectedElement?.tag || "project",
+                        group: scope === "component" ? "component" : "workspace",
+                        label: selectedElement?.text || "project",
+                        html: selectedElement ? `<${selectedElement.tag}> ${selectedElement.text || ""}` : undefined,
+                    },
+                };
+                const proposal = await (0, agent_1.proposeChanges)(cwd, input, config, signal, (message, detail, level) => {
+                    onProgress(level === "error" ? "error" : "working", message, detail);
+                }, (agentPrompt) => {
+                    onPrompt({ question: agentPrompt.message, options: agentPrompt.options });
+                });
+                if (signal.aborted) {
+                    return { ok: false, error: "Operation was cancelled" };
+                }
+                if (!proposal.changes.length) {
+                    return { ok: true, summary: proposal.summary, changes: [], thinking: [], totalThinkingTimeMs: Date.now() - startTime };
+                }
+                onProgress("review", "Implementation complete", `${proposal.changes.length} files modified`);
+                return {
+                    ok: true,
+                    summary: proposal.summary,
+                    changes: proposal.changes,
+                    thinking: [
+                        { title: "Analyzed project structure", detail: `Detected ${framework} framework` },
+                        { title: `${cliLabel(config.provider)} generated a proposal`, detail: `${proposal.changes.length} file(s)` },
+                    ],
+                    totalThinkingTimeMs: Date.now() - startTime,
+                };
+            }
+            catch (err) {
+                if (signal.aborted) {
+                    return { ok: false, error: "Operation was cancelled" };
+                }
+                return await gatewayFallback(err instanceof Error ? err.message : `${cliLabel(config.provider)} failed`);
+            }
+        }
+        if (useCliAgent && isConversational) {
+            try {
+                const reply = await generateConversationalReply(cwd, prompt, framework, config, signal, messages, pageContext, (message, detail) => onProgress("thinking", message, detail), taskId, (agentPrompt) => onPrompt({ question: agentPrompt.message, options: agentPrompt.options }));
+                return {
+                    ok: true,
+                    summary: reply,
+                    changes: [],
+                    thinking: [],
+                    totalThinkingTimeMs: Date.now() - startTime,
+                };
+            }
+            catch (err) {
+                if (signal.aborted) {
+                    return { ok: false, error: "Operation was cancelled" };
+                }
+                return await gatewayFallback(err instanceof Error ? err.message : `${cliLabel(config.provider)} failed`);
+            }
+        }
+        if (shouldRouteToServer) {
+            return await callGateway({ provider: config.provider, model: config.model });
         }
         // If it's a conversational prompt and server was not used or failed, respond directly without modifying files!
         if (isConversational) {
-            const reply = await generateConversationalReply(prompt, framework, config, signal, messages, pageContext);
+            const reply = await generateConversationalReply(cwd, prompt, framework, config, signal, messages, pageContext);
             return {
                 ok: true,
                 summary: reply,
@@ -578,11 +703,11 @@ Existing code context (study these carefully before writing — especially the C
 ${fileSnippets || "No existing components found. Create appropriate files in src/."}
 
 Analyze the existing code structure, then respond in JSON:`;
-    if (!config.apiKey && config.provider !== "ollama") {
+    if (!config.apiKey && config.provider !== "ollama" && !isCliProvider(config.provider)) {
         return {
             steps: [],
             isConversational: true,
-            reply: "Please configure an AI provider API key or model in settings to enable code generation.",
+            reply: "I couldn't generate code because no AI provider is available. Run `lasso auth login` to use your Lasso account, or add an AI provider API key in settings, then try again.",
         };
     }
     try {

package/dist/cli/project.js

@@ -211,7 +211,9 @@ async function initProject(cwd, fileEnv, forcedDomain) {
     try {
         let response;
         try {
-            response = await (0, auth_1.fetchWithTimeout)(`${collabApi}/collab/projects/register`, {
+            // Registration lives on the Lasso API (api.lasso.byorello.space): it owns
+            // the workspace's project records and stores the local `.lasso` domain.
+            response = await (0, auth_1.fetchWithTimeout)(`${serverUrl}/api/v1/collab/projects/register`, {
                 method: "POST",
                 headers: {
                     "content-type": "application/json",
@@ -221,8 +223,8 @@ async function initProject(cwd, fileEnv, forcedDomain) {
             });
         }
         catch {
-            // Fall back to main server if collab server unreachable
-            response = await (0, auth_1.fetchWithTimeout)(`${serverUrl}/api/v1/collab/projects/register`, {
+            // Fall back to the realtime/collab server if the API is unreachable
+            response = await (0, auth_1.fetchWithTimeout)(`${collabApi}/collab/projects/register`, {
                 method: "POST",
                 headers: {
                     "content-type": "application/json",
@@ -239,7 +241,8 @@ async function initProject(cwd, fileEnv, forcedDomain) {
             if (response.status === 403) {
                 return { ok: false, error: "Your API key does not have access to a Lasso workspace." };
             }
-            return { ok: false, error: body.message || `Project registration failed (${response.status}).` };
+            const detail = body.message || body.error;
+            return { ok: false, error: detail ? `${detail} (HTTP ${response.status})` : `Project registration failed (${response.status}).` };
         }
         const projectId = body.project?.id || existingConfig?.id;
         if (!projectId) {
@@ -264,7 +267,7 @@ async function initProject(cwd, fileEnv, forcedDomain) {
     }
     catch (error) {
         const reason = error instanceof Error ? error.message : "request failed";
-        return { ok: false, error: `Could not reach the Lasso realtime server at ${realtimeUrl} (${reason}). Is it running?` };
+        return { ok: false, error: `Could not reach the Lasso API at ${serverUrl} (${reason}). Is it running?` };
     }
 }
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lasso-ai/cli",
-  "version": "1.2.4",
+  "version": "1.2.6",
   "description": "Select any part of your running app, describe a change, and let AI edit the real source code.",
   "type": "commonjs",
   "main": "dist/cli/index.js",