fix: prevent context blowup by redacting thinking blocks between rounds

Thinking blocks from previous tool rounds (10-50KB each) were
accumulating in conversation history, causing Azure AI Foundry to hang
after 4+ rounds. Redact thinking content before each API call while
preserving required block structure. Also adds per-round timeout safety
net and status indicators between rounds.

js/cf-webapp/src/app/api/observability/chat/route.ts

@@ -16,6 +16,7 @@ interface ChatMessage {
 
 const MAX_TOOL_ROUNDS = 15
 const KEEPALIVE_INTERVAL_MS = 15_000
+const ROUND_TIMEOUT_MS = 3 * 60_000 // 3 minutes per API round
 
 const TOOL_DISPLAY_NAMES: Record<string, string> = {
   get_original_code: "Reading original code",
@@ -183,15 +184,35 @@ export async function POST(request: NextRequest): Promise<Response> {
         // Unified agent loop — each iteration either processes tool calls or
         // extracts the final text. Uses stream()+finalMessage() to avoid the
         // SDK's non-streaming timeout (max_tokens 32k estimates >10min).
-        // No separate "final streaming call" — when the loop gets a non-tool
-        // response, it extracts text directly, saving an extra API round-trip.
+        // Each round has a timeout to catch silent connection drops from the
+        // Azure AI Foundry proxy. Thinking blocks from older rounds are
+        // redacted (emptied) before each call to keep context size manageable.
         let toolRounds = 0
         while (toolRounds <= MAX_TOOL_ROUNDS) {
+          enqueue(`data: ${JSON.stringify({ type: "status", message: toolRounds === 0 ? "Thinking…" : "Analyzing…" })}\n\n`)
+
+          // Redact thinking content from previous assistant messages to prevent
+          // context blowup. The API requires the block structure but allows
+          // empty content. Each round's thinking can be 10-50KB.
+          for (const msg of conversationMessages) {
+            if (msg.role !== "assistant" || !Array.isArray(msg.content)) continue
+            for (const block of msg.content) {
+              if ((block as { type: string }).type === "thinking") {
+                (block as { thinking: string }).thinking = ""
+              }
+            }
+          }
+
           const messageStream = client.messages.stream(baseParams(systemPrompt, conversationMessages))
-          const response = await messageStream.finalMessage()
+          const timeout = setTimeout(() => messageStream.abort(), ROUND_TIMEOUT_MS)
+          let response: Anthropic.Message
+          try {
+            response = await messageStream.finalMessage()
+          } finally {
+            clearTimeout(timeout)
+          }
 
           if (response.stop_reason !== "tool_use") {
-            // Final response — extract text blocks and send to client
             for (const block of response.content) {
               if (block.type === "text") {
                 enqueue(`data: ${JSON.stringify({ type: "text", text: block.text })}\n\n`)

js/cf-webapp/src/app/observability/components/timeline-chat.tsx

@@ -34,6 +34,7 @@ export const TimelineChat = memo(function TimelineChat({
   const [isStreaming, setIsStreaming] = useState(false)
   const [completedRounds, setCompletedRounds] = useState<ToolStep[][]>([])
   const [activeSteps, setActiveSteps] = useState<ToolStep[]>([])
+  const [statusMessage, setStatusMessage] = useState<string | null>(null)
   const messagesEndRef = useRef<HTMLDivElement>(null)
   const inputRef = useRef<HTMLTextAreaElement>(null)
   const abortRef = useRef<AbortController | null>(null)
@@ -44,7 +45,7 @@ export const TimelineChat = memo(function TimelineChat({
 
   useEffect(() => {
     scrollToBottom()
-  }, [messages, completedRounds, activeSteps, scrollToBottom])
+  }, [messages, completedRounds, activeSteps, statusMessage, scrollToBottom])
 
   useEffect(() => {
     if (isOpen) {
@@ -104,8 +105,14 @@ export const TimelineChat = memo(function TimelineChat({
           try {
             const parsed = JSON.parse(data)
 
+            if (parsed.type === "status") {
+              setStatusMessage(parsed.message ?? null)
+              continue
+            }
+
             // Handle typed events (new protocol)
             if (parsed.type === "tool_start") {
+              setStatusMessage(null)
               setActiveSteps((prev) => {
                 // If all previous steps are done, this is a new round — commit previous steps
                 if (prev.length > 0 && prev.every((s) => s.status === "done")) {
@@ -142,6 +149,7 @@ export const TimelineChat = memo(function TimelineChat({
             // Handle text — both new {type: "text", text} and old {text} formats
             const textContent = parsed.type === "text" ? parsed.text : parsed.text
             if (textContent) {
+              setStatusMessage(null)
               setMessages((prev) => {
                 const updated = [...prev]
                 const last = updated[updated.length - 1]
@@ -185,6 +193,7 @@ export const TimelineChat = memo(function TimelineChat({
         }
         return []
       })
+      setStatusMessage(null)
       setIsStreaming(false)
       abortRef.current = null
     }
@@ -271,6 +280,13 @@ export const TimelineChat = memo(function TimelineChat({
 
         {activeSteps.length > 0 && <ToolRoundBubble steps={activeSteps} isActive />}
 
+        {statusMessage && (
+          <div className="flex items-center gap-2 text-xs text-zinc-400 dark:text-zinc-500 px-2">
+            <Loader2 className="h-3 w-3 animate-spin" />
+            <span>{statusMessage}</span>
+          </div>
+        )}
+
         <div ref={messagesEndRef} />
       </div>