v0.5.0: Binary-Free Mode - No OpenCode binary required

✨ Major Features:
- Native session management without OpenCode binary
- Provider routing: OpenCode Zen (free), Qwen OAuth, Z.AI
- Streaming chat with tool execution loop
- Mode detection API (/api/meta/mode)
- MCP integration fix (resolved infinite loading)
- NomadArch Native option in UI with comparison info

🆓 Free Models (No API Key):
- GPT-5 Nano (400K context)
- Grok Code Fast 1 (256K context)
- GLM-4.7 (205K context)
- Doubao Seed Code (256K context)
- Big Pickle (200K context)

📦 New Files:
- session-store.ts: Native session persistence
- native-sessions.ts: REST API for sessions
- lite-mode.ts: UI mode detection client
- native-sessions.ts (UI): SolidJS store

🔧 Updated:
- All installers: Optional binary download
- All launchers: Mode detection display
- Binary selector: Added NomadArch Native option
- README: Binary-Free Mode documentation

packages/ui/src/stores/session-actions.ts

@@ -1,5 +1,7 @@
+import { untrack, batch } from "solid-js"
+import { addDebugLog } from "../components/debug-overlay"
 import { resolvePastedPlaceholders } from "../lib/prompt-placeholders"
-import { instances } from "./instances"
+import { instances, activeInstanceId } from "./instances"
 import { addTaskMessage } from "./task-actions"
 
 import { addRecentModelPreference, setAgentModelPreference, getAgentModelPreference } from "./preferences"
@@ -36,7 +38,8 @@ const COMPACTION_ATTEMPT_TTL_MS = 60_000
 const COMPACTION_SUMMARY_MAX_CHARS = 4000
 const STREAM_TIMEOUT_MS = 120_000
 const OPENCODE_ZEN_OFFLINE_STORAGE_KEY = "opencode-zen-offline-models"
-const BUILD_PREVIEW_EVENT = "opencode:build-preview"
+export const BUILD_PREVIEW_EVENT = "opencode:build-preview"
+export const FILE_CHANGE_EVENT = "opencode:workspace-files-changed"
 
 function markOpencodeZenModelOffline(modelId: string): void {
   if (typeof window === "undefined" || !modelId) return
@@ -234,6 +237,8 @@ async function checkTokenBudgetBeforeSend(
 type ExternalChatMessage = { role: "user" | "assistant" | "system"; content: string }
 
 const MAX_ATTACHMENT_CHARS = 8000
+const MAX_CONTEXT_MESSAGES = 100
+const MAX_MESSAGES_FOR_YIELD = 50
 
 function shouldForceEnglish(prompt: string): boolean {
   const text = prompt.trim()
@@ -270,6 +275,12 @@ function clampText(value: string, maxChars: number): string {
   return `${value.slice(0, Math.max(0, maxChars - 3))}...`
 }
 
+async function yieldIfNeeded(index: number): Promise<void> {
+  if (index > 0 && index % MAX_MESSAGES_FOR_YIELD === 0) {
+    await new Promise(resolve => setTimeout(resolve, 0))
+  }
+}
+
 async function buildSkillsSystemInstruction(instanceId: string, sessionId: string): Promise<string | undefined> {
   const session = sessions().get(instanceId)?.get(sessionId)
   const selected = session?.skills ?? []
@@ -290,17 +301,42 @@ async function buildSkillsSystemInstruction(instanceId: string, sessionId: strin
   return `You have access to the following skills. Follow their instructions when relevant.\n\n${payload}`
 }
 
+async function buildFileSystemContext(instanceId: string): Promise<string | undefined> {
+  try {
+    const files = await serverApi.listWorkspaceFiles(instanceId)
+    if (!files || files.length === 0) return undefined
+
+    // Sort directories first
+    const sorted = files.sort((a: any, b: any) => {
+      const aDir = a.isDirectory || a.type === "directory"
+      const bDir = b.isDirectory || b.type === "directory"
+      if (aDir === bDir) return (a.name || "").localeCompare(b.name || "")
+      return aDir ? -1 : 1
+    })
+
+    const list = sorted.map((f: any) => {
+      const isDir = f.isDirectory || f.type === "directory"
+      return isDir ? `${f.name}/` : f.name
+    }).join("\n")
+
+    return `## Project Context\nCurrent Workspace Directory:\n\`\`\`\n${list}\n\`\`\`\nYou are an expert software architect working in this project. Use standard tools to explore further.`
+  } catch (error) {
+    return undefined
+  }
+}
+
 async function mergeSystemInstructions(
   instanceId: string,
   sessionId: string,
   prompt: string,
 ): Promise<string | undefined> {
-  const [languageSystem, skillsSystem] = await Promise.all([
+  const [languageSystem, skillsSystem, projectContext] = await Promise.all([
     Promise.resolve(buildLanguageSystemInstruction(prompt)),
     buildSkillsSystemInstruction(instanceId, sessionId),
+    buildFileSystemContext(instanceId),
   ])
   const sshInstruction = buildSshPasswordInstruction(prompt)
-  const sections = [languageSystem, skillsSystem, sshInstruction].filter(Boolean) as string[]
+  const sections = [projectContext, languageSystem, skillsSystem, sshInstruction].filter(Boolean) as string[]
   if (sections.length === 0) return undefined
   return sections.join("\n\n")
 }
@@ -346,32 +382,40 @@ function extractPlainTextFromParts(
   return segments.join("\n").trim()
 }
 
-function buildExternalChatMessages(
+async function buildExternalChatMessages(
   instanceId: string,
   sessionId: string,
   systemMessage?: string,
-): ExternalChatMessage[] {
-  const store = messageStoreBus.getOrCreate(instanceId)
-  const messageIds = store.getSessionMessageIds(sessionId)
-  const messages: ExternalChatMessage[] = []
+): Promise<ExternalChatMessage[]> {
+  return untrack(async () => {
+    const store = messageStoreBus.getOrCreate(instanceId)
+    const messageIds = store.getSessionMessageIds(sessionId)
+    const messages: ExternalChatMessage[] = []
 
-  if (systemMessage) {
-    messages.push({ role: "system", content: systemMessage })
-  }
+    if (systemMessage) {
+      messages.push({ role: "system", content: systemMessage })
+    }
 
-  for (const messageId of messageIds) {
-    const record = store.getMessage(messageId)
-    if (!record) continue
-    const { orderedParts } = buildRecordDisplayData(instanceId, record)
-    const content = extractPlainTextFromParts(orderedParts as Array<{ type?: string; text?: unknown; filename?: string }>)
-    if (!content) continue
-    messages.push({
-      role: record.role === "assistant" ? "assistant" : "user",
-      content,
-    })
-  }
+    const limitedMessageIds = messageIds.length > MAX_CONTEXT_MESSAGES
+      ? messageIds.slice(-MAX_CONTEXT_MESSAGES)
+      : messageIds
 
-  return messages
+    for (let i = 0; i < limitedMessageIds.length; i++) {
+      const messageId = limitedMessageIds[i]
+      await yieldIfNeeded(i)
+      const record = store.getMessage(messageId)
+      if (!record) continue
+      const { orderedParts } = buildRecordDisplayData(instanceId, record)
+      const content = extractPlainTextFromParts(orderedParts as Array<{ type?: string; text?: unknown; filename?: string }>)
+      if (!content) continue
+      messages.push({
+        role: record.role === "assistant" ? "assistant" : "user",
+        content,
+      })
+    }
+
+    return messages
+  })
 }
 
 function decodeAttachmentData(data: Uint8Array): string {
@@ -391,7 +435,7 @@ async function buildExternalChatMessagesWithAttachments(
   systemMessage: string | undefined,
   attachments: Array<{ filename?: string; source?: any; mediaType?: string }>,
 ): Promise<ExternalChatMessage[]> {
-  const baseMessages = buildExternalChatMessages(instanceId, sessionId, systemMessage)
+  const baseMessages = await buildExternalChatMessages(instanceId, sessionId, systemMessage)
   if (!attachments || attachments.length === 0) {
     return baseMessages
   }
@@ -455,6 +499,8 @@ async function readSseStream(
   resetIdleTimer()
 
   try {
+    let chunkCount = 0
+    let lastYieldTime = performance.now()
     while (!shouldStop) {
       const { done, value } = await reader.read()
       if (done) break
@@ -473,9 +519,21 @@ async function readSseStream(
           break
         }
         onData(data)
+        chunkCount++
+      }
+      // Throttle UI updates: yield control if time elapsed > 16ms to prevent frame drops
+      const now = performance.now()
+      if (now - lastYieldTime > 16) {
+        addDebugLog(`Yielding after ${Math.round(now - lastYieldTime)}ms (chunks: ${chunkCount})`, "info")
+        lastYieldTime = now
+        if ('requestIdleCallback' in window) {
+          await new Promise<void>(resolve => {
+            requestIdleCallback(() => resolve(), { timeout: 16 })
+          })
+        } else {
+          await new Promise<void>(resolve => setTimeout(resolve, 0))
+        }
       }
-      // Yield to main thread periodically to prevent UI freeze during rapid streaming
-      await new Promise<void>(resolve => setTimeout(resolve, 0))
     }
     if (timedOut) {
       throw new Error("Stream timed out")
@@ -499,6 +557,10 @@ async function streamOllamaChat(
   const controller = new AbortController()
   const timeoutId = setTimeout(() => controller.abort(), STREAM_TIMEOUT_MS)
 
+  // Get workspace path for tool execution
+  const instance = instances().get(instanceId)
+  const workspacePath = instance?.folder || ""
+
   const response = await fetch("/api/ollama/chat", {
     method: "POST",
     headers: { "Content-Type": "application/json" },
@@ -507,6 +569,8 @@ async function streamOllamaChat(
       model: modelId,
       messages,
       stream: true,
+      workspacePath,
+      enableTools: true,
     }),
   })
 
@@ -516,54 +580,105 @@ async function streamOllamaChat(
   }
 
   const store = messageStoreBus.getOrCreate(instanceId)
+  store.beginStreamingUpdate()
   let fullText = ""
+  let lastUpdateAt = 0
 
   try {
     await readSseStream(response, (data) => {
       try {
         const chunk = JSON.parse(data)
-        // Check for error response from server
-        if (chunk?.error) {
-          throw new Error(chunk.error)
+        if (chunk?.error) throw new Error(chunk.error)
+
+        // Handle tool execution results (special events from backend)
+        if (chunk?.type === "tool_result") {
+          const toolResult = `\n\n✅ **Tool Executed:** ${chunk.content}\n\n`
+          fullText += toolResult
+          store.applyPartUpdate({
+            messageId: assistantMessageId,
+            part: { id: assistantPartId, type: "text", text: fullText } as any,
+          })
+
+          // Dispatch file change event to refresh sidebar
+          if (typeof window !== "undefined") {
+            console.log(`[EVENT] Dispatching FILE_CHANGE_EVENT for ${instanceId}`);
+            window.dispatchEvent(new CustomEvent(FILE_CHANGE_EVENT, { detail: { instanceId } }))
+          }
+
+          // Auto-trigger preview for HTML file writes
+          const content = chunk.content || ""
+          if (content.includes("Successfully wrote") &&
+            (content.includes(".html") || content.includes("index.") || content.includes(".htm"))) {
+            if (typeof window !== "undefined") {
+              const htmlMatch = content.match(/to\s+([^\s]+\.html?)/)
+              if (htmlMatch) {
+                const relativePath = htmlMatch[1]
+                const origin = typeof window !== "undefined" ? window.location.origin : "http://localhost:3000"
+                const apiOrigin = origin.replace(":3000", ":9898")
+                const previewUrl = `${apiOrigin}/api/workspaces/${instanceId}/serve/${relativePath}`
+
+                console.log(`[EVENT] Auto-preview triggered for ${previewUrl}`);
+                window.dispatchEvent(new CustomEvent(BUILD_PREVIEW_EVENT, {
+                  detail: { url: previewUrl, instanceId }
+                }))
+              }
+            }
+          }
+
+          return
         }
+
         const delta = chunk?.message?.content
         if (typeof delta !== "string" || delta.length === 0) return
         fullText += delta
-        store.applyPartUpdate({
-          messageId: assistantMessageId,
-          part: { id: assistantPartId, type: "text", text: fullText } as any,
-        })
+
+        const now = Date.now()
+        if (now - lastUpdateAt > 150) { // Limit to ~7 updates per second
+          lastUpdateAt = now
+          store.applyPartUpdate({
+            messageId: assistantMessageId,
+            part: { id: assistantPartId, type: "text", text: fullText } as any,
+          })
+        }
       } catch (e) {
         if (e instanceof Error) throw e
-        // Ignore malformed chunks
       }
     })
+
+    // Always apply final text update
+    store.applyPartUpdate({
+      messageId: assistantMessageId,
+      part: { id: assistantPartId, type: "text", text: fullText } as any,
+    })
   } finally {
     clearTimeout(timeoutId)
+    store.endStreamingUpdate()
   }
 
-  store.upsertMessage({
-    id: assistantMessageId,
-    sessionId,
-    role: "assistant",
-    status: "complete",
-    updatedAt: Date.now(),
-    isEphemeral: false,
-  })
-  store.setMessageInfo(assistantMessageId, {
-    id: assistantMessageId,
-    role: "assistant",
-    providerID: providerId,
-    modelID: modelId,
-    time: { created: store.getMessageInfo(assistantMessageId)?.time?.created ?? Date.now(), completed: Date.now() },
-  } as any)
-  store.upsertMessage({
-    id: messageId,
-    sessionId,
-    role: "user",
-    status: "sent",
-    updatedAt: Date.now(),
-    isEphemeral: false,
+  batch(() => {
+    store.upsertMessage({
+      id: assistantMessageId,
+      sessionId,
+      role: "assistant",
+      status: "complete",
+      updatedAt: Date.now(),
+      isEphemeral: false,
+    })
+    store.setMessageInfo(assistantMessageId, {
+      id: assistantMessageId,
+      role: "assistant",
+      providerID: providerId,
+      modelID: modelId,
+      time: { created: store.getMessageInfo(assistantMessageId)?.time?.created ?? Date.now(), completed: Date.now() },
+    } as any)
+    store.upsertMessage({
+      id: messageId,
+      sessionId,
+      role: "user",
+      status: "sent",
+      updatedAt: Date.now(),
+      isEphemeral: false,
+    })
   })
 }
 
@@ -582,6 +697,10 @@ async function streamQwenChat(
   const controller = new AbortController()
   const timeoutId = setTimeout(() => controller.abort(), STREAM_TIMEOUT_MS)
 
+  // Get workspace path for tool execution
+  const instance = instances().get(instanceId)
+  const workspacePath = instance?.folder || ""
+
   const response = await fetch("/api/qwen/chat", {
     method: "POST",
     headers: {
@@ -594,6 +713,8 @@ async function streamQwenChat(
       messages,
       stream: true,
       resource_url: resourceUrl,
+      workspacePath,
+      enableTools: true,
     }),
   })
 
@@ -603,27 +724,86 @@ async function streamQwenChat(
   }
 
   const store = messageStoreBus.getOrCreate(instanceId)
+  store.beginStreamingUpdate()
   let fullText = ""
+  let lastUpdateAt = 0
 
   try {
     await readSseStream(response, (data) => {
       try {
         const chunk = JSON.parse(data)
+
+        // Handle tool execution results
+        if (chunk?.type === "tool_result") {
+          const toolResult = `\n\n✅ **Tool Executed:** ${chunk.content}\n\n`
+          fullText += toolResult
+          store.applyPartUpdate({
+            messageId: assistantMessageId,
+            part: { id: assistantPartId, type: "text", text: fullText } as any,
+          })
+
+          // Dispatch file change event to refresh sidebar
+          if (typeof window !== "undefined") {
+            console.log(`[Qwen] Dispatching FILE_CHANGE_EVENT for ${instanceId}`);
+            console.log(`[EVENT] Dispatching FILE_CHANGE_EVENT for ${instanceId}`);
+            window.dispatchEvent(new CustomEvent(FILE_CHANGE_EVENT, { detail: { instanceId } }));
+
+            // Double-tap refresh after 1s to catch FS latency
+            setTimeout(() => {
+              window.dispatchEvent(new CustomEvent(FILE_CHANGE_EVENT, { detail: { instanceId } }));
+            }, 1000);
+          }
+
+          // Auto-trigger preview for HTML file writes
+          const content = chunk.content || ""
+          if (content.includes("Successfully wrote") &&
+            (content.includes(".html") || content.includes("index.") || content.includes(".htm"))) {
+            if (typeof window !== "undefined") {
+              const htmlMatch = content.match(/to\s+([^\s]+\.html?)/)
+              if (htmlMatch) {
+                const relativePath = htmlMatch[1]
+                const origin = typeof window !== "undefined" ? window.location.origin : "http://localhost:3000"
+                const apiOrigin = origin.replace(":3000", ":9898")
+                const previewUrl = `${apiOrigin}/api/workspaces/${instanceId}/serve/${relativePath}`
+
+                console.log(`[Qwen] Auto-preview triggered for ${relativePath}`);
+                console.log(`[EVENT] Auto-preview triggered for ${previewUrl}`);
+                window.dispatchEvent(new CustomEvent(BUILD_PREVIEW_EVENT, {
+                  detail: { url: previewUrl, instanceId }
+                }))
+              }
+            }
+          }
+          return
+        }
+
         const delta =
           chunk?.choices?.[0]?.delta?.content ??
           chunk?.choices?.[0]?.message?.content
         if (typeof delta !== "string" || delta.length === 0) return
         fullText += delta
-        store.applyPartUpdate({
-          messageId: assistantMessageId,
-          part: { id: assistantPartId, type: "text", text: fullText } as any,
-        })
+
+        const now = Date.now()
+        if (now - lastUpdateAt > 40) { // Limit to ~25 updates per second
+          lastUpdateAt = now
+          store.applyPartUpdate({
+            messageId: assistantMessageId,
+            part: { id: assistantPartId, type: "text", text: fullText } as any,
+          })
+        }
       } catch {
         // Ignore malformed chunks
       }
     })
+
+    // Always apply final text update
+    store.applyPartUpdate({
+      messageId: assistantMessageId,
+      part: { id: assistantPartId, type: "text", text: fullText } as any,
+    })
   } finally {
     clearTimeout(timeoutId)
+    store.endStreamingUpdate()
   }
 
   store.upsertMessage({
@@ -664,6 +844,10 @@ async function streamOpenCodeZenChat(
   const controller = new AbortController()
   const timeoutId = setTimeout(() => controller.abort(), STREAM_TIMEOUT_MS)
 
+  // Get workspace path for tool execution
+  const instance = instances().get(instanceId)
+  const workspacePath = instance?.folder || ""
+
   const response = await fetch("/api/opencode-zen/chat", {
     method: "POST",
     headers: { "Content-Type": "application/json" },
@@ -672,6 +856,8 @@ async function streamOpenCodeZenChat(
       model: modelId,
       messages,
       stream: true,
+      workspacePath,
+      enableTools: true,
     }),
   })
 
@@ -681,7 +867,9 @@ async function streamOpenCodeZenChat(
   }
 
   const store = messageStoreBus.getOrCreate(instanceId)
+  store.beginStreamingUpdate()
   let fullText = ""
+  let lastUpdateAt = 0
 
   try {
     await readSseStream(response, (data) => {
@@ -690,23 +878,78 @@ async function streamOpenCodeZenChat(
         if (chunk?.error) {
           throw new Error(typeof chunk.error === "string" ? chunk.error : "OpenCode Zen streaming error")
         }
+
+        // Handle tool execution results (special events from backend)
+        if (chunk?.type === "tool_result") {
+          const toolResult = `\n\n✅ **Tool Executed:** ${chunk.content}\n\n`
+          fullText += toolResult
+          store.applyPartUpdate({
+            messageId: assistantMessageId,
+            part: { id: assistantPartId, type: "text", text: fullText } as any,
+          })
+
+          // Dispatch file change event to refresh sidebar
+          if (typeof window !== "undefined") {
+            console.log(`[Ollama] Dispatching FILE_CHANGE_EVENT for ${instanceId}`);
+            console.log(`[EVENT] Dispatching FILE_CHANGE_EVENT for ${instanceId}`);
+            window.dispatchEvent(new CustomEvent(FILE_CHANGE_EVENT, { detail: { instanceId } }))
+          }
+
+          // Auto-trigger preview for HTML file writes
+          const content = chunk.content || ""
+          if (content.includes("Successfully wrote") &&
+            (content.includes(".html") || content.includes("index.") || content.includes(".htm"))) {
+            if (typeof window !== "undefined") {
+              const htmlMatch = content.match(/to\s+([^\s]+\.html?)/)
+              if (htmlMatch) {
+                const relativePath = htmlMatch[1]
+                // USE PROXY URL instead of file:// to avoid "Not allowed to load local resource"
+                // The backend (port 9898) serves workspace files via /api/workspaces/:id/serve
+                const origin = typeof window !== "undefined" ? window.location.origin : "http://localhost:3000"
+                const apiOrigin = origin.replace(":3000", ":9898") // Fallback assumption
+                const previewUrl = `${apiOrigin}/api/workspaces/${instanceId}/serve/${relativePath}`
+
+                console.log(`[Ollama] Auto-preview triggered for ${relativePath}`);
+                console.log(`[EVENT] Auto-preview triggered for ${previewUrl}`);
+                window.dispatchEvent(new CustomEvent(BUILD_PREVIEW_EVENT, {
+                  detail: { url: previewUrl, instanceId }
+                }))
+              }
+            }
+          }
+
+          return
+        }
+
         const delta =
           chunk?.choices?.[0]?.delta?.content ??
           chunk?.choices?.[0]?.message?.content
         if (typeof delta !== "string" || delta.length === 0) return
         fullText += delta
-        store.applyPartUpdate({
-          messageId: assistantMessageId,
-          part: { id: assistantPartId, type: "text", text: fullText } as any,
-        })
+
+        const now = Date.now()
+        if (now - lastUpdateAt > 40) { // Limit to ~25 updates per second
+          lastUpdateAt = now
+          store.applyPartUpdate({
+            messageId: assistantMessageId,
+            part: { id: assistantPartId, type: "text", text: fullText } as any,
+          })
+        }
       } catch (error) {
         if (error instanceof Error) {
           throw error
         }
       }
     })
+
+    // Always apply final text update
+    store.applyPartUpdate({
+      messageId: assistantMessageId,
+      part: { id: assistantPartId, type: "text", text: fullText } as any,
+    })
   } finally {
     clearTimeout(timeoutId)
+    store.endStreamingUpdate()
   }
 
 
@@ -748,6 +991,10 @@ async function streamZAIChat(
   const controller = new AbortController()
   const timeoutId = setTimeout(() => controller.abort(), STREAM_TIMEOUT_MS)
 
+  // Get workspace path for tool execution
+  const instance = instances().get(instanceId)
+  const workspacePath = instance?.folder || ""
+
   const response = await fetch("/api/zai/chat", {
     method: "POST",
     headers: { "Content-Type": "application/json" },
@@ -756,6 +1003,8 @@ async function streamZAIChat(
       model: modelId,
       messages,
       stream: true,
+      workspacePath,
+      enableTools: true,
     }),
   })
 
@@ -765,32 +1014,81 @@ async function streamZAIChat(
   }
 
   const store = messageStoreBus.getOrCreate(instanceId)
+  store.beginStreamingUpdate()
   let fullText = ""
+  let lastUpdateAt = 0
 
   try {
     await readSseStream(response, (data) => {
       try {
         const chunk = JSON.parse(data)
-        // Check for error response from server
-        if (chunk?.error) {
-          throw new Error(chunk.error)
+        if (chunk?.error) throw new Error(chunk.error)
+
+        // Handle tool execution results (special events from backend)
+        if (chunk?.type === "tool_result") {
+          const toolResult = `\n\n✅ **Tool Executed:** ${chunk.content}\n\n`
+          fullText += toolResult
+          store.applyPartUpdate({
+            messageId: assistantMessageId,
+            part: { id: assistantPartId, type: "text", text: fullText } as any,
+          })
+
+          // Dispatch file change event to refresh sidebar
+          if (typeof window !== "undefined") {
+            console.log(`[EVENT] Dispatching FILE_CHANGE_EVENT for ${instanceId}`);
+            window.dispatchEvent(new CustomEvent(FILE_CHANGE_EVENT, { detail: { instanceId } }))
+          }
+
+          // Auto-trigger preview for HTML file writes
+          const content = chunk.content || ""
+          if (content.includes("Successfully wrote") &&
+            (content.includes(".html") || content.includes("index.") || content.includes(".htm"))) {
+            if (typeof window !== "undefined") {
+              const htmlMatch = content.match(/to\s+([^\s]+\.html?)/)
+              if (htmlMatch) {
+                const relativePath = htmlMatch[1]
+                const origin = typeof window !== "undefined" ? window.location.origin : "http://localhost:3000"
+                const apiOrigin = origin.replace(":3000", ":9898")
+                const previewUrl = `${apiOrigin}/api/workspaces/${instanceId}/serve/${relativePath}`
+
+                console.log(`[EVENT] Auto-preview triggered for ${previewUrl}`);
+                window.dispatchEvent(new CustomEvent(BUILD_PREVIEW_EVENT, {
+                  detail: { url: previewUrl, instanceId }
+                }))
+              }
+            }
+          }
+
+          return
         }
+
         const delta =
           chunk?.choices?.[0]?.delta?.content ??
           chunk?.choices?.[0]?.message?.content
         if (typeof delta !== "string" || delta.length === 0) return
         fullText += delta
-        store.applyPartUpdate({
-          messageId: assistantMessageId,
-          part: { id: assistantPartId, type: "text", text: fullText } as any,
-        })
+
+        const now = Date.now()
+        if (now - lastUpdateAt > 40) { // Limit to ~25 updates per second
+          lastUpdateAt = now
+          store.applyPartUpdate({
+            messageId: assistantMessageId,
+            part: { id: assistantPartId, type: "text", text: fullText } as any,
+          })
+        }
       } catch (e) {
         if (e instanceof Error) throw e
-        // Ignore malformed chunks
       }
     })
+
+    // Always apply final text update
+    store.applyPartUpdate({
+      messageId: assistantMessageId,
+      part: { id: assistantPartId, type: "text", text: fullText } as any,
+    })
   } finally {
     clearTimeout(timeoutId)
+    store.endStreamingUpdate()
   }
 
   store.upsertMessage({
@@ -941,15 +1239,17 @@ async function sendMessage(
 
   log.info("sendMessage: upserting optimistic message", { messageId, sessionId, taskId });
 
-  store.upsertMessage({
-    id: messageId,
-    sessionId,
-    role: "user",
-    status: "sending",
-    parts: optimisticParts,
-    createdAt,
-    updatedAt: createdAt,
-    isEphemeral: true,
+  untrack(() => {
+    store.upsertMessage({
+      id: messageId,
+      sessionId,
+      role: "user",
+      status: "sending",
+      parts: optimisticParts,
+      createdAt,
+      updatedAt: createdAt,
+      isEphemeral: true,
+    })
   })
 
   withSession(instanceId, sessionId, () => {
@@ -957,47 +1257,62 @@ async function sendMessage(
   })
 
   const providerId = effectiveModel.providerId
-  const systemMessage = await mergeSystemInstructions(instanceId, sessionId, prompt)
+  const tPre1 = performance.now()
+  const systemMessage = await untrack(() => mergeSystemInstructions(instanceId, sessionId, prompt))
+  const tPre2 = performance.now()
+  if (tPre2 - tPre1 > 10) {
+    addDebugLog(`Merge System Instructions: ${Math.round(tPre2 - tPre1)}ms`, "warn")
+  }
+
   if (providerId === "ollama-cloud" || providerId === "qwen-oauth" || providerId === "opencode-zen" || providerId === "zai") {
     const store = messageStoreBus.getOrCreate(instanceId)
     const now = Date.now()
     const assistantMessageId = createId("msg")
     const assistantPartId = createId("part")
+
+    const tMsg1 = performance.now()
     const externalMessages = await buildExternalChatMessagesWithAttachments(
       instanceId,
       sessionId,
       systemMessage,
       attachments,
     )
+    const tMsg2 = performance.now()
+    if (tMsg2 - tMsg1 > 10) {
+      addDebugLog(`Build External Messages: ${Math.round(tMsg2 - tMsg1)}ms`, "warn")
+    }
 
-    store.upsertMessage({
-      id: assistantMessageId,
-      sessionId,
-      role: "assistant",
-      status: "streaming",
-      parts: [{ id: assistantPartId, type: "text", text: "" } as any],
-      createdAt: now,
-      updatedAt: now,
-      isEphemeral: true,
-    })
-    store.setMessageInfo(assistantMessageId, {
-      id: assistantMessageId,
-      role: "assistant",
-      providerID: effectiveModel.providerId,
-      modelID: effectiveModel.modelId,
-      time: { created: now, completed: 0 },
-    } as any)
-    store.upsertMessage({
-      id: messageId,
-      sessionId,
-      role: "user",
-      status: "sent",
-      updatedAt: now,
-      isEphemeral: false,
+    untrack(() => {
+      store.upsertMessage({
+        id: assistantMessageId,
+        sessionId,
+        role: "assistant",
+        status: "streaming",
+        parts: [{ id: assistantPartId, type: "text", text: "" } as any],
+        createdAt: now,
+        updatedAt: now,
+        isEphemeral: true,
+      })
+      store.setMessageInfo(assistantMessageId, {
+        id: assistantMessageId,
+        role: "assistant",
+        providerID: effectiveModel.providerId,
+        modelID: effectiveModel.modelId,
+        time: { created: now, completed: 0 },
+      } as any)
+      store.upsertMessage({
+        id: messageId,
+        sessionId,
+        role: "user",
+        status: "sent",
+        updatedAt: now,
+        isEphemeral: false,
+      })
     })
 
     try {
       if (providerId === "ollama-cloud") {
+        const tStream1 = performance.now()
         await streamOllamaChat(
           instanceId,
           sessionId,
@@ -1008,6 +1323,8 @@ async function sendMessage(
           assistantMessageId,
           assistantPartId,
         )
+        const tStream2 = performance.now()
+        addDebugLog(`Stream Complete: ${Math.round(tStream2 - tStream1)}ms`, "info")
       } else if (providerId === "opencode-zen") {
         await streamOpenCodeZenChat(
           instanceId,
@@ -1370,7 +1687,7 @@ async function compactSession(instanceId: string, sessionId: string): Promise<Co
       const tasksCopy = session.tasks.map((task) => ({ ...task }))
       withSession(instanceId, compactedSession.id, (nextSession) => {
         nextSession.tasks = tasksCopy
-        nextSession.activeTaskId = undefined
+        nextSession.activeTaskId = session.activeTaskId
       })
     }
 
@@ -1632,6 +1949,48 @@ async function forkSession(instanceId: string, sessionId: string): Promise<strin
   }
 }
 
+// Forcefully reset streaming state to unlock UI if stuck
+function forceReset() {
+  const store = messageStoreBus.getOrCreate(activeInstanceId() || "")
+  if (!store) return
+
+  // Reset streaming count forcefully
+  // We don't have direct access to set count to 0, so we call end enough times
+  // or we assume we can just ignore it for now, but really we should expose a reset method.
+  // For now, let's just log and clear pending parts.
+  store.setState("pendingParts", {})
+
+  // If we could access the store's internal streaming count setter that would be better.
+  // Since we added `isStreaming` and `endStreamingUpdate` to store interface,
+  // we can just call end multiple times if we suspect it's stuck > 0
+  let safety = 0
+  while (store.state.streamingUpdateCount > 0 && safety < 100) {
+    store.endStreamingUpdate()
+    safety++
+  }
+
+  // Also reset message statuses
+  try {
+    const messages = store.state.messages;
+    Object.values(messages).forEach(msg => {
+      if (msg.status === "streaming" || msg.status === "sending") {
+        store.upsertMessage({
+          id: msg.id,
+          sessionId: msg.sessionId,
+          role: msg.role,
+          status: "interrupted",
+          updatedAt: Date.now(),
+          isEphemeral: msg.isEphemeral,
+        })
+      }
+    })
+  } catch (e) {
+    console.error("Error updating message status during reset", e)
+  }
+
+  addDebugLog("Force Reset Triggered: Cleared streaming state & statuses", "warn")
+}
+
 export {
   abortSession,
   compactSession,
@@ -1644,4 +2003,5 @@ export {
   updateSessionAgent,
   updateSessionModel,
   updateSessionModelForSession,
+  forceReset, // Add to exports
 }