@mindstudio-ai/remy 0.1.53 → 0.1.55

package/dist/headless.js

@@ -6,6 +6,7 @@ var __export = (target, all) => {
 
 // src/headless.ts
 import { createInterface } from "readline";
+import { writeFileSync } from "fs";
 
 // src/logger.ts
 import fs from "fs";
@@ -406,6 +407,10 @@ ${isLspConfigured() ? `<typescript_lsp>
 
 {{static/instructions.md}}
 
+<conversation_summaries>
+Your conversation history may include <prior_conversation_summary> blocks in the user's messages. These are automated summaries of earlier messages that have been compacted to save context space. The user does not see this summary, they see the full conversation history in their UI. Treat the summary as ground truth for what happened before, but do not reference it directly to the user ("as mentioned in the summary..."). Just continue naturally as if you remember the prior work.
+</conversation_summaries>
+
 <!-- cache_breakpoint -->
 
 <project_onboarding>
@@ -636,6 +641,181 @@ async function generateBackgroundAck(params) {
   }
 }
 
+// src/compaction/index.ts
+var log4 = createLogger("compaction");
+var CONVERSATION_SUMMARY_PROMPT = readAsset("compaction", "conversation.md");
+var SUBAGENT_SUMMARY_PROMPT = readAsset("compaction", "subagent.md");
+var SUMMARIZABLE_SUBAGENTS = ["visualDesignExpert", "productVision"];
+async function compactConversation(state, apiConfig) {
+  const insertionIndex = state.messages.length;
+  const summaries = [];
+  const tasks = [];
+  const conversationMessages = getConversationMessagesForSummary(
+    state.messages,
+    insertionIndex
+  );
+  if (conversationMessages.length > 0) {
+    tasks.push(
+      generateSummary(
+        apiConfig,
+        "conversation",
+        CONVERSATION_SUMMARY_PROMPT,
+        conversationMessages
+      ).then((text) => {
+        if (text) {
+          summaries.push({ name: "conversation", text });
+        }
+      })
+    );
+  }
+  for (const name of SUMMARIZABLE_SUBAGENTS) {
+    const subagentMessages = getSubAgentMessagesForSummary(
+      state.messages,
+      name,
+      insertionIndex
+    );
+    if (subagentMessages.length > 0) {
+      tasks.push(
+        generateSummary(
+          apiConfig,
+          name,
+          SUBAGENT_SUMMARY_PROMPT,
+          subagentMessages
+        ).then((text) => {
+          if (text) {
+            summaries.push({ name, text });
+          }
+        })
+      );
+    }
+  }
+  await Promise.all(tasks);
+  const checkpointMessages = summaries.map((s) => ({
+    role: "user",
+    hidden: true,
+    content: [
+      {
+        type: "summary",
+        name: s.name,
+        text: s.text,
+        startedAt: Date.now()
+      }
+    ]
+  }));
+  if (checkpointMessages.length > 0) {
+    state.messages.splice(insertionIndex, 0, ...checkpointMessages);
+  }
+  log4.info("Compaction complete", {
+    summaries: summaries.length,
+    insertionIndex,
+    messagesAfter: state.messages.length - insertionIndex - checkpointMessages.length
+  });
+}
+function getConversationMessagesForSummary(messages, endIndex) {
+  let startIdx = 0;
+  for (let i = endIndex - 1; i >= 0; i--) {
+    const msg = messages[i];
+    if (!Array.isArray(msg.content)) {
+      continue;
+    }
+    for (const block of msg.content) {
+      if (block.type === "summary" && block.name === "conversation") {
+        startIdx = i + 1;
+        break;
+      }
+    }
+    if (startIdx > 0) {
+      break;
+    }
+  }
+  return messages.slice(startIdx, endIndex);
+}
+function getSubAgentMessagesForSummary(messages, subAgentName, endIndex) {
+  let checkpointIdx = -1;
+  for (let i = endIndex - 1; i >= 0; i--) {
+    const msg = messages[i];
+    if (!Array.isArray(msg.content)) {
+      continue;
+    }
+    for (const block of msg.content) {
+      if (block.type === "summary" && block.name === subAgentName) {
+        checkpointIdx = i;
+        break;
+      }
+    }
+    if (checkpointIdx !== -1) {
+      break;
+    }
+  }
+  const startIdx = checkpointIdx !== -1 ? checkpointIdx + 1 : 0;
+  const collected = [];
+  for (let i = startIdx; i < endIndex; i++) {
+    const msg = messages[i];
+    if (msg.role !== "assistant" || !Array.isArray(msg.content)) {
+      continue;
+    }
+    for (const block of msg.content) {
+      if (block.type === "tool" && block.name === subAgentName && block.subAgentMessages?.length) {
+        collected.push(...block.subAgentMessages);
+      }
+    }
+  }
+  return collected;
+}
+function serializeForSummary(messages) {
+  return messages.map((msg) => {
+    if (typeof msg.content === "string") {
+      return `[${msg.role}]: ${msg.content}`;
+    }
+    if (!Array.isArray(msg.content)) {
+      return `[${msg.role}]: (empty)`;
+    }
+    const blocks = msg.content;
+    const parts = [];
+    for (const block of blocks) {
+      if (block.type === "text") {
+        parts.push(block.text);
+      } else if (block.type === "tool") {
+        parts.push(
+          `[tool: ${block.name}(${JSON.stringify(block.input).slice(0, 200)})] \u2192 ${(block.result ?? "").slice(0, 500)}`
+        );
+      }
+    }
+    return `[${msg.role}]: ${parts.join("\n")}`;
+  }).join("\n\n");
+}
+async function generateSummary(apiConfig, name, systemPrompt, messagesToSummarize) {
+  const serialized = serializeForSummary(messagesToSummarize);
+  if (!serialized.trim()) {
+    return null;
+  }
+  log4.info("Generating summary", {
+    name,
+    messageCount: messagesToSummarize.length
+  });
+  let summaryText = "";
+  for await (const event of streamChat({
+    ...apiConfig,
+    subAgentId: "conversationSummarizer",
+    system: systemPrompt,
+    messages: [{ role: "user", content: serialized }],
+    tools: []
+  })) {
+    if (event.type === "text") {
+      summaryText += event.text;
+    } else if (event.type === "error") {
+      log4.error("Summary generation failed", { name, error: event.error });
+      return null;
+    }
+  }
+  if (!summaryText.trim()) {
+    log4.warn("Empty summary generated", { name });
+    return null;
+  }
+  log4.info("Summary generated", { name, summaryLength: summaryText.length });
+  return summaryText.trim();
+}
+
 // src/tools/spec/readSpec.ts
 import fs5 from "fs/promises";
 
@@ -1417,40 +1597,6 @@ var askMindStudioSdkTool = {
   }
 };
 
-// src/tools/common/fetchUrl.ts
-var fetchUrlTool = {
-  definition: {
-    name: "scapeWebUrl",
-    description: "Scrape the content of a web page. Returns the HTML of the page as markdown text. Optionally capture a screenshot if you need see the visual design. Use this when you need to fetch or analyze content from a website",
-    inputSchema: {
-      type: "object",
-      properties: {
-        url: {
-          type: "string",
-          description: "The URL to fetch."
-        },
-        screenshot: {
-          type: "boolean",
-          description: "Capture a screenshot of the page in addition to the text content. Adds latency; only use when you need to see the visual design."
-        }
-      },
-      required: ["url"]
-    }
-  },
-  async execute(input, context) {
-    const url = input.url;
-    const screenshot = input.screenshot;
-    const pageOptions = { onlyMainContent: true };
-    if (screenshot) {
-      pageOptions.screenshot = true;
-    }
-    return runCli(
-      `mindstudio scrape-url --url ${JSON.stringify(url)} --page-options ${JSON.stringify(JSON.stringify(pageOptions))} --no-meta`,
-      { onLog: context?.onLog }
-    );
-  }
-};
-
 // src/tools/common/searchGoogle.ts
 var searchGoogleTool = {
   definition: {
@@ -2179,21 +2325,28 @@ async function analyzeImage(params) {
 var SCREENSHOT_ANALYSIS_PROMPT = "Describe everything visible on screen from top to bottom \u2014 every element, its position, its size relative to the viewport, its colors, its content. Be comprehensive, thorough, and spatial. After the inventory, note anything that looks visually broken (overlapping elements, clipped text, misaligned components). Respond only with your analysis as Markdown and absolutely no other text. Do not use emojis - use unicode if you need symbols.";
 async function captureAndAnalyzeScreenshot(promptOrOptions) {
   let prompt;
+  let existingUrl;
   let onLog;
   if (typeof promptOrOptions === "object" && promptOrOptions !== null) {
     prompt = promptOrOptions.prompt;
+    existingUrl = promptOrOptions.imageUrl;
     onLog = promptOrOptions.onLog;
   } else {
     prompt = promptOrOptions;
   }
-  const ssResult = await sidecarRequest("/screenshot-full-page", void 0, {
-    timeout: 12e4
-  });
-  const url = ssResult?.url || ssResult?.screenshotUrl;
-  if (!url) {
-    throw new Error(
-      `No URL in sidecar response. The browser may not be ready yet. Response: ${JSON.stringify(ssResult)}`
-    );
+  let url;
+  if (existingUrl) {
+    url = existingUrl;
+  } else {
+    const ssResult = await sidecarRequest("/screenshot-full-page", void 0, {
+      timeout: 12e4
+    });
+    url = ssResult?.url || ssResult?.screenshotUrl;
+    if (!url) {
+      throw new Error(
+        `No URL in sidecar response. The browser may not be ready yet. Response: ${JSON.stringify(ssResult)}`
+      );
+    }
   }
   if (prompt === false) {
     return url;
@@ -2211,19 +2364,30 @@ async function captureAndAnalyzeScreenshot(promptOrOptions) {
 var screenshotTool = {
   definition: {
     name: "screenshot",
-    description: "Capture a full-height screenshot of the app preview and get a description of what's on screen. Optionally provide a specific question about what you're looking for..",
+    description: "Capture a full-height screenshot of the app preview and get a description of what's on screen. Provides static image analysis only, will not capture animations or video. Optionally provide specific questions about what you're looking for. Use a bulleted list to ask many questions at once. To ask additional questions about a screenshot you have already captured, pass its URL as imageUrl to skip recapture.",
     inputSchema: {
       type: "object",
       properties: {
         prompt: {
           type: "string",
           description: "Optional question about the screenshot. If omitted, returns a general description of what's visible."
+        },
+        imageUrl: {
+          type: "string",
+          description: "URL of an existing screenshot to analyze instead of capturing a new one. Use this for additional questions about a previous screenshot."
         }
       }
     }
   },
   async execute(input, context) {
     try {
+      if (input.imageUrl) {
+        return await captureAndAnalyzeScreenshot({
+          prompt: input.prompt,
+          imageUrl: input.imageUrl,
+          onLog: context?.onLog
+        });
+      }
       return await captureAndAnalyzeScreenshot({
         prompt: input.prompt,
         onLog: context?.onLog
@@ -2289,8 +2453,51 @@ function startStatusWatcher(config) {
 }
 
 // src/subagents/common/cleanMessages.ts
+function findLastSummaryCheckpoint(messages, name) {
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const msg = messages[i];
+    if (!Array.isArray(msg.content)) {
+      continue;
+    }
+    for (const block of msg.content) {
+      if (block.type === "summary" && block.name === name) {
+        return i;
+      }
+    }
+  }
+  return -1;
+}
 function cleanMessagesForApi(messages) {
-  return messages.map((msg) => {
+  const checkpointIdx = findLastSummaryCheckpoint(messages, "conversation");
+  let startIdx = 0;
+  const prefix = [];
+  if (checkpointIdx !== -1) {
+    const checkpointMsg = messages[checkpointIdx];
+    const blocks = checkpointMsg.content;
+    const summaryBlock = blocks.find(
+      (b) => b.type === "summary" && b.name === "conversation"
+    );
+    if (summaryBlock && summaryBlock.type === "summary") {
+      prefix.push({
+        role: "user",
+        content: `<conversation_summary>
+${summaryBlock.text}
+</conversation_summary>`,
+        hidden: true
+      });
+    }
+    startIdx = checkpointIdx + 1;
+  }
+  const messagesToProcess = messages.slice(startIdx);
+  const cleaned = messagesToProcess.filter((msg) => {
+    if (Array.isArray(msg.content)) {
+      const blocks = msg.content;
+      if (blocks.some((b) => b.type === "summary")) {
+        return false;
+      }
+    }
+    return true;
+  }).map((msg) => {
     if (msg.role === "user" && typeof msg.content === "string" && msg.content.startsWith("@@automated::")) {
       return {
         ...msg,
@@ -2306,25 +2513,26 @@ function cleanMessagesForApi(messages) {
     const thinking = blocks.filter(
       (b) => b.type === "thinking"
     ).map((b) => ({ thinking: b.thinking, signature: b.signature }));
-    const cleaned = {
+    const cleaned2 = {
       role: msg.role,
       content: text
     };
     if (toolCalls.length > 0) {
-      cleaned.toolCalls = toolCalls;
+      cleaned2.toolCalls = toolCalls;
     }
     if (thinking.length > 0) {
-      cleaned.thinking = thinking;
+      cleaned2.thinking = thinking;
     }
     if (msg.hidden) {
-      cleaned.hidden = true;
+      cleaned2.hidden = true;
     }
-    return cleaned;
+    return cleaned2;
   });
+  return [...prefix, ...cleaned];
 }
 
 // src/subagents/runner.ts
-var log4 = createLogger("sub-agent");
+var log5 = createLogger("sub-agent");
 async function runSubAgent(config) {
   const {
     system,
@@ -2341,6 +2549,7 @@ async function runSubAgent(config) {
     resolveExternalTool,
     toolRegistry,
     requestId,
+    history,
     background,
     onBackgroundComplete
   } = config;
@@ -2348,13 +2557,16 @@ async function runSubAgent(config) {
   const signal = background ? bgAbort.signal : parentSignal;
   const agentName = subAgentId || "sub-agent";
   const runStart = Date.now();
-  log4.info("Sub-agent started", { requestId, parentToolId, agentName });
+  log5.info("Sub-agent started", { requestId, parentToolId, agentName });
   const emit2 = (e) => {
     onEvent({ ...e, parentToolId });
   };
   let turns = 0;
   const run = async () => {
-    const messages = [{ role: "user", content: task }];
+    const messages = [
+      ...history ?? [],
+      { role: "user", content: task }
+    ];
     function getPartialText(blocks) {
       return blocks.filter((b) => b.type === "text").map((b) => b.text).join("");
     }
@@ -2501,7 +2713,7 @@ Current date/time: ${(/* @__PURE__ */ new Date()).toISOString().replace("T", " "
         const text = getPartialText(contentBlocks);
         return { text, messages };
       }
-      log4.info("Tools executing", {
+      log5.info("Tools executing", {
         requestId,
         parentToolId,
         count: toolCalls.length,
@@ -2572,7 +2784,7 @@ Current date/time: ${(/* @__PURE__ */ new Date()).toISOString().replace("T", " "
           run2(tc.input);
           const r = await resultPromise;
           toolRegistry?.unregister(tc.id);
-          log4.info("Tool completed", {
+          log5.info("Tool completed", {
             requestId,
             parentToolId,
             toolCallId: tc.id,
@@ -2613,7 +2825,7 @@ Current date/time: ${(/* @__PURE__ */ new Date()).toISOString().replace("T", " "
   const wrapRun = async () => {
     try {
       const result = await run();
-      log4.info("Sub-agent complete", {
+      log5.info("Sub-agent complete", {
         requestId,
         parentToolId,
         agentName,
@@ -2622,7 +2834,7 @@ Current date/time: ${(/* @__PURE__ */ new Date()).toISOString().replace("T", " "
       });
       return result;
     } catch (err) {
-      log4.warn("Sub-agent error", {
+      log5.warn("Sub-agent error", {
         requestId,
         parentToolId,
         agentName,
@@ -2634,7 +2846,7 @@ Current date/time: ${(/* @__PURE__ */ new Date()).toISOString().replace("T", " "
   if (!background) {
     return wrapRun();
   }
-  log4.info("Sub-agent backgrounded", { requestId, parentToolId, agentName });
+  log5.info("Sub-agent backgrounded", { requestId, parentToolId, agentName });
   toolRegistry?.register({
     id: parentToolId,
     name: agentName,
@@ -2781,7 +2993,7 @@ ${appSpec}
 }
 
 // src/subagents/browserAutomation/index.ts
-var log5 = createLogger("browser-automation");
+var log6 = createLogger("browser-automation");
 var browserAutomationTool = {
   definition: {
     name: "runAutomatedBrowserTest",
@@ -2876,7 +3088,7 @@ var browserAutomationTool = {
                   }
                 }
               } catch {
-                log5.debug("Failed to parse batch analysis result", {
+                log6.debug("Failed to parse batch analysis result", {
                   batchResult
                 });
               }
@@ -2902,7 +3114,7 @@ __export(searchGoogle_exports, {
 });
 var definition = {
   name: "searchGoogle",
-  description: 'Search Google for web results. Reserch modern design trends in industries or verticals, "best [domain] apps 2026", ui patterns, or find something specific if the the user has an explicit reference. Prioritize authoritative sources like Figma and other design leaders, avoid random blog spam. Pick one or more URLs from the results and then use `fetchUrl` to get their text content.',
+  description: 'Search Google for web results. Reserch modern design trends in industries or verticals, "best [domain] apps 2026", ui patterns, or find something specific if the the user has an explicit reference. Searching for and reading case studies is a great way to get information and context about a project\'s domain. Prioritize authoritative sources like Figma and other design leaders, avoid random blog spam. Pick one or more URLs from the results and then use `scrapeWebUrl` to get their text content.',
   inputSchema: {
     type: "object",
     properties: {
@@ -2921,14 +3133,14 @@ async function execute(input, onLog) {
   );
 }
 
-// src/subagents/designExpert/tools/fetchUrl.ts
-var fetchUrl_exports = {};
-__export(fetchUrl_exports, {
+// src/subagents/designExpert/tools/scrapeWebUrl.ts
+var scrapeWebUrl_exports = {};
+__export(scrapeWebUrl_exports, {
   definition: () => definition2,
   execute: () => execute2
 });
 var definition2 = {
-  name: "fetchUrl",
+  name: "scrapeWebUrl",
   description: "Fetch the content of a web page as markdown. Use when reading sites from search results or specific things the user wants to incorporate.",
   inputSchema: {
     type: "object",
@@ -2983,7 +3195,7 @@ Respond only with your analysis as Markdown and absolutely no other text. Do not
 `;
 var definition3 = {
   name: "analyzeDesign",
-  description: "Analyze the visual design of a website or image URL. Websites are automatically screenshotted first. If no prompt is provided, performs a full design reference analysis (mood, color, typography, layout, distinctiveness). Provide a custom prompt to ask a specific design question instead.",
+  description: "Analyze the visual design of a website or image URL. Websites are automatically screenshotted first. Provides static image analysis only, will not capture animations or video. If no prompt is provided, performs a full design reference analysis (mood, color, typography, layout, distinctiveness). Provide a custom prompt to ask a specific design question instead. Use a bulleted list to ask many questions at once.",
   inputSchema: {
     type: "object",
     properties: {
@@ -3028,10 +3240,10 @@ __export(analyzeImage_exports, {
   definition: () => definition4,
   execute: () => execute4
 });
-var DEFAULT_PROMPT = "Describe everything visible in this image \u2014 every element, its position, its size relative to the frame, its colors, its content. Be comprhensive, thorough and spatial. After the inventory, note anything that looks visually broken (overlapping elements, clipped text, misaligned components). Respond only with your analysis as Markdown and absolutely no other text. Do not use emojis - use unicode if you need symbols.";
+var DEFAULT_PROMPT = "Describe everything visible in this image \u2014 every element, its position, its size relative to the frame, its colors, its content. Be comprehensive, thorough and spatial. After the inventory, note anything that looks visually broken (overlapping elements, clipped text, misaligned components). Respond only with your analysis as Markdown and absolutely no other text. Do not use emojis - use unicode if you need symbols.";
 var definition4 = {
   name: "analyzeImage",
-  description: "Analyze an image by URL using a vision model. Returns an objective description of what is visible \u2014 shapes, colors, layout, text, artifacts. Use for factual inventory of image contents, not for subjective design judgment - the vision model providing the analysis has no sense of design. You are the design expert - use the analysis tool for factual inventory, then apply your own expertise for quality and suitability assessments.",
+  description: "Analyze an image by URL using a vision model. Provides static image analysis only, will not capture animations or video. Returns an objective description of what is visible \u2014 shapes, colors, layout, text, artifacts. Use for factual inventory of image contents, not for subjective design judgment - the vision model providing the analysis has no sense of design. You are the design expert - use the analysis tool for factual inventory, then apply your own expertise for quality and suitability assessments. Optionally provide specific questions about what you're looking for. Use a bulleted list to ask many questions at once. If you are analyzing a screenshot of the app preview, you can reuse the same screenshot URL multiple times to ask multiple questions.",
   inputSchema: {
     type: "object",
     properties: {
@@ -3330,7 +3542,7 @@ async function execute7(input, onLog) {
 // src/subagents/designExpert/tools/index.ts
 var tools = {
   searchGoogle: searchGoogle_exports,
-  fetchUrl: fetchUrl_exports,
+  scrapeWebUrl: scrapeWebUrl_exports,
   analyzeDesign: analyzeDesign_exports,
   analyzeImage: analyzeImage_exports,
   screenshot: screenshot_exports,
@@ -3587,6 +3799,50 @@ ${specContext}`;
   return prompt;
 }
 
+// src/subagents/common/history.ts
+function getSubAgentHistory(messages, subAgentName) {
+  let checkpointIdx = -1;
+  let summaryText = "";
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const msg = messages[i];
+    if (!Array.isArray(msg.content)) {
+      continue;
+    }
+    for (const block of msg.content) {
+      if (block.type === "summary" && block.name === subAgentName) {
+        checkpointIdx = i;
+        summaryText = block.text;
+        break;
+      }
+    }
+    if (checkpointIdx !== -1) {
+      break;
+    }
+  }
+  const history = [];
+  if (checkpointIdx !== -1 && summaryText) {
+    history.push({
+      role: "user",
+      content: `<prior_conversation_summary>
+${summaryText}
+</prior_conversation_summary>`
+    });
+  }
+  const startIdx = checkpointIdx !== -1 ? checkpointIdx + 1 : 0;
+  for (let i = startIdx; i < messages.length; i++) {
+    const msg = messages[i];
+    if (msg.role !== "assistant" || !Array.isArray(msg.content)) {
+      continue;
+    }
+    for (const block of msg.content) {
+      if (block.type === "tool" && block.name === subAgentName && block.subAgentMessages?.length) {
+        history.push(...block.subAgentMessages);
+      }
+    }
+  }
+  return history;
+}
+
 // src/subagents/designExpert/index.ts
 var DESCRIPTION = `
 Visual design expert. Describe the situation and what you need \u2014 the agent decides what to deliver. It reads the spec files automatically. Include relevant user requirements and context it can't get from the spec, but do not list specific deliverables or tell it how to do its job. Do not suggest implementation details or ideas - only relay what is needed.
@@ -3614,9 +3870,11 @@ var designExpertTool = {
     if (!context) {
       return "Error: visual design expert requires execution context";
     }
+    const history = context.conversationMessages ? getSubAgentHistory(context.conversationMessages, "visualDesignExpert") : [];
     const result = await runSubAgent({
       system: getDesignExpertPrompt(),
       task: input.task,
+      history: history.length > 0 ? history : void 0,
       tools: DESIGN_EXPERT_TOOLS,
       externalTools: /* @__PURE__ */ new Set(),
       executeTool: (name, input2, toolCallId, onLog) => executeDesignExpertTool(name, input2, context, toolCallId, onLog),
@@ -3912,9 +4170,11 @@ var productVisionTool = {
     if (!context) {
       return "Error: product vision requires execution context";
     }
+    const history = context.conversationMessages ? getSubAgentHistory(context.conversationMessages, "productVision") : [];
     const result = await runSubAgent({
       system: getProductVisionPrompt(),
       task: input.task,
+      history: history.length > 0 ? history : void 0,
       tools: VISION_TOOLS,
       externalTools: /* @__PURE__ */ new Set(),
       executeTool: executeVisionTool,
@@ -3999,7 +4259,7 @@ var SANITY_CHECK_TOOLS = [
     }
   },
   {
-    name: "fetchUrl",
+    name: "scrapeWebUrl",
     description: "Fetch a web page as markdown. Use to read package docs, changelogs, npm pages.",
     inputSchema: {
       type: "object",
@@ -4082,6 +4342,40 @@ var codeSanityCheckTool = {
   }
 };
 
+// src/tools/common/scrapeWebUrl.ts
+var scrapeWebUrlTool = {
+  definition: {
+    name: "scrapeWebUrl",
+    description: "Scrape the content of a web page. Returns the HTML of the page as markdown text. Optionally capture a screenshot if you need see the visual design. Use this when you need to fetch or analyze content from a website",
+    inputSchema: {
+      type: "object",
+      properties: {
+        url: {
+          type: "string",
+          description: "The URL to fetch."
+        },
+        screenshot: {
+          type: "boolean",
+          description: "Capture a screenshot of the page in addition to the text content. Adds latency; only use when you need to see the visual design."
+        }
+      },
+      required: ["url"]
+    }
+  },
+  async execute(input, context) {
+    const url = input.url;
+    const screenshot = input.screenshot;
+    const pageOptions = { onlyMainContent: true };
+    if (screenshot) {
+      pageOptions.screenshot = true;
+    }
+    return runCli(
+      `mindstudio scrape-url --url ${JSON.stringify(url)} --page-options ${JSON.stringify(JSON.stringify(pageOptions))} --no-meta`,
+      { onLog: context?.onLog }
+    );
+  }
+};
+
 // src/tools/index.ts
 function getSpecTools() {
   return [readSpecTool, writeSpecTool, editSpecTool, listSpecFilesTool];
@@ -4112,7 +4406,7 @@ function getCommonTools() {
     promptUserTool,
     confirmDestructiveActionTool,
     askMindStudioSdkTool,
-    fetchUrlTool,
+    scrapeWebUrlTool,
     searchGoogleTool,
     setProjectMetadataTool,
     designExpertTool,
@@ -4165,7 +4459,7 @@ function executeTool(name, input, context) {
 
 // src/session.ts
 import fs17 from "fs";
-var log6 = createLogger("session");
+var log7 = createLogger("session");
 var SESSION_FILE = ".remy-session.json";
 function loadSession(state) {
   try {
@@ -4173,7 +4467,7 @@ function loadSession(state) {
     const data = JSON.parse(raw);
     if (Array.isArray(data.messages) && data.messages.length > 0) {
       state.messages = sanitizeMessages(data.messages);
-      log6.info("Session loaded", { messageCount: state.messages.length });
+      log7.info("Session loaded", { messageCount: state.messages.length });
       return true;
     }
   } catch {
@@ -4223,9 +4517,9 @@ function saveSession(state) {
       JSON.stringify({ messages: state.messages }, null, 2),
       "utf-8"
     );
-    log6.info("Session saved", { messageCount: state.messages.length });
+    log7.info("Session saved", { messageCount: state.messages.length });
   } catch (err) {
-    log6.warn("Session save failed", { error: err.message });
+    log7.warn("Session save failed", { error: err.message });
   }
 }
 function clearSession(state) {
@@ -4426,7 +4720,7 @@ function friendlyError(raw) {
 }
 
 // src/agent.ts
-var log7 = createLogger("agent");
+var log8 = createLogger("agent");
 function getTextContent(blocks) {
   return blocks.filter((b) => b.type === "text").map((b) => b.text).join("");
 }
@@ -4469,7 +4763,7 @@ async function runTurn(params) {
     onBackgroundComplete
   } = params;
   const tools2 = getToolDefinitions(onboardingState);
-  log7.info("Turn started", {
+  log8.info("Turn started", {
     requestId,
     model,
     toolCount: tools2.length,
@@ -4495,6 +4789,11 @@ async function runTurn(params) {
   let lastCompletedTools = "";
   let lastCompletedInput = "";
   let lastCompletedResult = "";
+  let turnInputTokens = 0;
+  let turnOutputTokens = 0;
+  let turnCacheCreation = 0;
+  let turnCacheRead = 0;
+  let turnLlmCalls = 0;
   while (true) {
     let getOrCreateAccumulator2 = function(id, name) {
       let acc = toolInputAccumulators.get(id);
@@ -4671,7 +4970,7 @@ async function runTurn(params) {
             const tool = getToolByName(event.name);
             const wasStreamed = acc?.started ?? false;
             const isInputStreaming = !!tool?.streaming?.partialInput;
-            log7.info("Tool received", {
+            log8.info("Tool received", {
               requestId,
               toolCallId: event.id,
               name: event.name
@@ -4688,6 +4987,11 @@ async function runTurn(params) {
           }
           case "done":
             stopReason = event.stopReason;
+            turnLlmCalls++;
+            turnInputTokens += event.usage.inputTokens;
+            turnOutputTokens += event.usage.outputTokens;
+            turnCacheCreation += event.usage.cacheCreationTokens ?? 0;
+            turnCacheRead += event.usage.cacheReadTokens ?? 0;
             break;
           case "error":
             onEvent({ type: "error", error: friendlyError(event.error) });
@@ -4725,10 +5029,19 @@ async function runTurn(params) {
     if (stopReason !== "tool_use" || toolCalls.length === 0) {
       statusWatcher.stop();
       saveSession(state);
-      onEvent({ type: "turn_done" });
+      onEvent({
+        type: "turn_done",
+        stats: {
+          inputTokens: turnInputTokens,
+          outputTokens: turnOutputTokens,
+          cacheCreationTokens: turnCacheCreation || void 0,
+          cacheReadTokens: turnCacheRead || void 0,
+          llmCalls: turnLlmCalls
+        }
+      });
       return;
     }
-    log7.info("Tools executing", {
+    log8.info("Tools executing", {
       requestId,
       count: toolCalls.length,
       tools: toolCalls.map((tc) => tc.name)
@@ -4772,7 +5085,7 @@ async function runTurn(params) {
             let result;
             if (EXTERNAL_TOOLS.has(tc.name) && resolveExternalTool) {
               saveSession(state);
-              log7.info("Waiting for external tool result", {
+              log8.info("Waiting for external tool result", {
                 requestId,
                 toolCallId: tc.id,
                 name: tc.name
@@ -4788,6 +5101,7 @@ async function runTurn(params) {
                 toolCallId: tc.id,
                 requestId,
                 subAgentMessages,
+                conversationMessages: state.messages,
                 toolRegistry,
                 onBackgroundComplete,
                 onLog: (line) => wrappedOnEvent({
@@ -4827,7 +5141,7 @@ async function runTurn(params) {
         if (!tc.input.background) {
           toolRegistry?.unregister(tc.id);
         }
-        log7.info("Tool completed", {
+        log8.info("Tool completed", {
           requestId,
           toolCallId: tc.id,
           name: tc.name,
@@ -4882,7 +5196,7 @@ async function runTurn(params) {
 }
 
 // src/toolRegistry.ts
-var log8 = createLogger("tool-registry");
+var log9 = createLogger("tool-registry");
 var ToolRegistry = class {
   entries = /* @__PURE__ */ new Map();
   onEvent;
@@ -4908,7 +5222,7 @@ var ToolRegistry = class {
     if (!entry) {
       return false;
     }
-    log8.info("Tool stopped", { toolCallId: id, name: entry.name, mode });
+    log9.info("Tool stopped", { toolCallId: id, name: entry.name, mode });
     entry.abortController.abort(mode);
     if (mode === "graceful") {
       const partial = entry.getPartialResult?.() ?? "";
@@ -4941,7 +5255,7 @@ ${partial}` : "[INTERRUPTED] Tool execution was stopped.";
     if (!entry) {
       return false;
     }
-    log8.info("Tool restarted", { toolCallId: id, name: entry.name });
+    log9.info("Tool restarted", { toolCallId: id, name: entry.name });
     entry.abortController.abort("restart");
     const newInput = patchedInput ? { ...entry.input, ...patchedInput } : entry.input;
     this.onEvent?.({
@@ -4985,7 +5299,7 @@ ${body}`;
 }
 
 // src/headless.ts
-var log9 = createLogger("headless");
+var log10 = createLogger("headless");
 function emit(event, data, requestId) {
   const payload = { event, ...data };
   if (requestId) {
@@ -5049,6 +5363,17 @@ async function startHeadless(opts = {}) {
   const pendingTools = /* @__PURE__ */ new Map();
   const earlyResults = /* @__PURE__ */ new Map();
   const toolRegistry = new ToolRegistry();
+  const sessionStats = {
+    messageCount: 0,
+    turns: 0,
+    totalInputTokens: 0,
+    totalOutputTokens: 0,
+    totalCacheCreationTokens: 0,
+    totalCacheReadTokens: 0,
+    lastContextSize: 0,
+    compactionInProgress: false,
+    updatedAt: 0
+  };
   const backgroundQueue = [];
   function flushBackgroundQueue() {
     if (backgroundQueue.length === 0) {
@@ -5066,22 +5391,32 @@ ${xmlParts}
 </background_results>`;
     handleMessage({ action: "message", text: message }, void 0);
   }
-  function onBackgroundComplete(toolCallId, name, result, subAgentMessages) {
-    for (const msg of state.messages) {
-      if (!Array.isArray(msg.content)) {
-        continue;
-      }
-      for (const block of msg.content) {
-        if (block.type === "tool" && block.id === toolCallId) {
-          block.backgroundResult = result;
-          block.completedAt = Date.now();
-          if (subAgentMessages) {
-            block.subAgentMessages = subAgentMessages;
+  const pendingBlockUpdates = [];
+  function applyPendingBlockUpdates() {
+    if (pendingBlockUpdates.length === 0) {
+      return;
+    }
+    const updates = pendingBlockUpdates.splice(0);
+    for (const update of updates) {
+      for (const msg of state.messages) {
+        if (!Array.isArray(msg.content)) {
+          continue;
+        }
+        for (const block of msg.content) {
+          if (block.type === "tool" && block.id === update.toolCallId) {
+            block.backgroundResult = update.result;
+            block.completedAt = Date.now();
+            if (update.subAgentMessages) {
+              block.subAgentMessages = update.subAgentMessages;
+            }
           }
         }
       }
     }
-    log9.info("Background complete", {
+  }
+  function onBackgroundComplete(toolCallId, name, result, subAgentMessages) {
+    pendingBlockUpdates.push({ toolCallId, result, subAgentMessages });
+    log10.info("Background complete", {
       toolCallId,
       name,
       requestId: currentRequestId
@@ -5099,6 +5434,7 @@ ${xmlParts}
       completedAt: Date.now()
     });
     if (!running) {
+      applyPendingBlockUpdates();
       flushBackgroundQueue();
     }
   }
@@ -5141,8 +5477,25 @@ ${xmlParts}
       // Terminal events — translate to `completed`
       case "turn_done":
         completedEmitted = true;
+        if (e.stats) {
+          sessionStats.turns++;
+          sessionStats.totalInputTokens += e.stats.inputTokens;
+          sessionStats.totalOutputTokens += e.stats.outputTokens;
+          sessionStats.totalCacheCreationTokens += e.stats.cacheCreationTokens ?? 0;
+          sessionStats.totalCacheReadTokens += e.stats.cacheReadTokens ?? 0;
+          sessionStats.lastContextSize = e.stats.inputTokens;
+        }
+        sessionStats.messageCount = state.messages.length;
+        sessionStats.updatedAt = Date.now();
+        try {
+          writeFileSync(".remy-stats.json", JSON.stringify(sessionStats));
+        } catch {
+        }
         emit("completed", { success: true }, rid);
-        setTimeout(() => flushBackgroundQueue(), 0);
+        setTimeout(() => {
+          applyPendingBlockUpdates();
+          flushBackgroundQueue();
+        }, 0);
         return;
       case "turn_cancelled":
         completedEmitted = true;
@@ -5331,7 +5684,7 @@ ${xmlParts}
           requestId
         );
       }
-      log9.info("Turn complete", {
+      log10.info("Turn complete", {
         requestId,
         durationMs: Date.now() - turnStart
       });
@@ -5340,7 +5693,7 @@ ${xmlParts}
         emit("error", { error: err.message }, requestId);
         emit("completed", { success: false, error: err.message }, requestId);
       }
-      log9.warn("Command failed", {
+      log10.warn("Command failed", {
         action: "message",
         requestId,
         error: err.message
@@ -5360,7 +5713,7 @@ ${xmlParts}
       return;
     }
     const { action, requestId } = parsed;
-    log9.info("Command received", { action, requestId });
+    log10.info("Command received", { action, requestId });
     if (action === "tool_result" && parsed.id) {
       const id = parsed.id;
       const result = parsed.result ?? "";
@@ -5416,6 +5769,39 @@ ${xmlParts}
       }
       return;
     }
+    if (action === "compact") {
+      sessionStats.compactionInProgress = true;
+      sessionStats.updatedAt = Date.now();
+      try {
+        writeFileSync(".remy-stats.json", JSON.stringify(sessionStats));
+      } catch {
+      }
+      compactConversation(state, config).then(() => {
+        saveSession(state);
+        emit("compaction_complete", {}, requestId);
+        emit("completed", { success: true }, requestId);
+      }).catch((err) => {
+        emit(
+          "compaction_complete",
+          { error: err.message || "Compaction failed" },
+          requestId
+        );
+        emit(
+          "completed",
+          { success: false, error: err.message || "Compaction failed" },
+          requestId
+        );
+      }).finally(() => {
+        sessionStats.compactionInProgress = false;
+        sessionStats.messageCount = state.messages.length;
+        sessionStats.updatedAt = Date.now();
+        try {
+          writeFileSync(".remy-stats.json", JSON.stringify(sessionStats));
+        } catch {
+        }
+      });
+      return;
+    }
     if (action === "message") {
       await handleMessage(parsed, requestId);
       return;